Stochastic gradient algorithms are often unstable when applied to functions that do not have Lipschitz-continuous and/or bounded gradients. Gradient clipping is a simple and effective technique to stabilize the training process for problems that are prone to the exploding gradient problem. Despite its widespread popularity, the convergence properties of the gradient clipping heuristic are poorly understood, especially for stochastic problems. This paper establishes both qualitative and quantitative convergence results of the clipped stochastic (sub)gradient method (SGD) for non-smooth convex functions with rapidly growing subgradients. Our analyses show that clipping enhances the stability of SGD and that the clipped SGD algorithm enjoys finite convergence rates in many cases. We also study the convergence of a clipped method with momentum, which includes clipped SGD as a special case, for weakly convex problems under standard assumptions. With a novel Lyapunov analysis, we show that the proposed method achieves the best-known rate for the considered class of problems, demonstrating the effectiveness of clipped methods also in this regime. Numerical results confirm our theoretical developments.

Adder neural network (AdderNet) is a new kind of deep model that replaces the original massive multiplications in convolutions by additions while preserving the high performance. Since the hardware complexity of additions is much lower than that of multiplications, the overall energy consumption is thus reduced significantly. To further optimize the hardware overhead of using AdderNet, this paper studies the winograd algorithm, which is a widely used fast algorithm for accelerating convolution and saving the computational costs. Unfortunately, the conventional Winograd algorithm cannot be directly applied to AdderNets since the distributive law in multiplication is not valid for the l1-norm. Therefore, we replace the element-wise multiplication in the Winograd equation by additions and then develop a new set of transform matrixes that can enhance the representation ability of output features to maintain the performance. Moreover, we propose the l2-to-l1 training strategy to mitigate the negative impacts caused by formal inconsistency. Experimental results on both FPGA and benchmarks show that the new method can further reduce the energy consumption without affecting the accuracy of the original AdderNet.

The Euclidean projection onto a convex set is an important problem that arises in numerous constrained optimization tasks. Unfortunately, in many cases, computing projections is computationally demanding. In this work, we focus on projection problems where the constraints are smooth and the number of constraints is significantly smaller than the dimension. The runtime of existing approaches to solving such problems is either cubic in the dimension or polynomial in the inverse of the target accuracy. Conversely, we propose a simple and efficient primal-dual approach, with a runtime that scales only linearly with the dimension, and only logarithmically in the inverse of the target accuracy. We empirically demonstrate its performance, and compare it with standard baselines.

Banks are required to analyse large transaction datasets as a part of the fight against financial crime. Today, this analysis is either performed manually by domain experts or using expensive feature engineering. Gradient flow analysis allows for basic representation learning as node potentials can be inferred directly from network transaction data. However, the gradient model has a fundamental limitation: it cannot represent all types of of network flows. Furthermore, standard methods for learning the gradient flow are not appropriate for flow signals that span multiple orders of magnitude and contain outliers, i.e. transaction data. In this work, the gradient model is extended to a gated version and we prove that it, unlike the gradient model, is a universal approximator for flows on graphs. To tackle the mentioned challenges of transaction data, we propose a multi-scale and outlier robust loss function based on the Student-t log-likelihood. Ethereum transaction data is used for evaluation and the gradient models outperform MLP models using hand-engineered and node2vec features in terms of relative error. These results extend to 60 synthetic datasets, with experiments also showing that the gated gradient model learns qualitative information about the underlying synthetic generative flow distributions.

We propose a hierarchical version of dual averaging for zeroth-order online non-convex optimization – i.e., learning processes where, at each stage, the optimizer is facing an unknown non-convex loss function and only receives the incurred loss as feedback. The proposed class of policies relies on the construction of an online model that aggregates loss information as it arrives, and it consists of two principal components: (a) a regularizer adapted to the Fisher information metric (as opposed to the metric norm of the ambient space); and (b) a principled exploration of the problem’s state space based on an adapted hierarchical schedule. This construction enables sharper control of the model’s bias and variance, and allows us to derive tight bounds for both the learner’s static and dynamic regret – i.e., the regret incurred against the best dynamic policy in hindsight over the horizon of play.

We address the problem of teaching a deep reinforcement learning (RL) agent to follow instructions in multi-task environments. Instructions are expressed in a well-known formal language – linear temporal logic (LTL) – and can specify a diversity of complex, temporally extended behaviours, including conditionals and alternative realizations. Our proposed learning approach exploits the compositional syntax and the semantics of LTL, enabling our RL agent to learn task-conditioned policies that generalize to new instructions, not observed during training. To reduce the overhead of learning LTL semantics, we introduce an environment-agnostic LTL pretraining scheme which improves sample-efficiency in downstream environments. Experiments on discrete and continuous domains target combinatorial task sets of up to $\sim10^{39}$ unique tasks and demonstrate the strength of our approach in learning to solve (unseen) tasks, given LTL instructions.

The lack of anisotropic kernels in graph neural networks (GNNs) strongly limits their expressiveness, contributing to well-known issues such as over-smoothing. To overcome this limitation, we propose the first globally consistent anisotropic kernels for GNNs, allowing for graph convolutions that are defined according to topologicaly-derived directional flows. First, by defining a vector field in the graph, we develop a method of applying directional derivatives and smoothing by projecting node-specific messages into the field. Then, we propose the use of the Laplacian eigenvectors as such vector field. We show that the method generalizes CNNs on an $n$-dimensional grid and is provably more discriminative than standard GNNs regarding the Weisfeiler-Lehman 1-WL test. We evaluate our method on different standard benchmarks and see a relative error reduction of 8% on the CIFAR10 graph dataset and 11% to 32% on the molecular ZINC dataset, and a relative increase in precision of 1.6% on the MolPCBA dataset. An important outcome of this work is that it enables graph networks to embed directions in an unsupervised way, thus allowing a better representation of the anisotropic features in different physical or biological problems.

Traditional supervised learning aims to train a classifier in the closed-set world, where training and test samples share the same label space. In this paper, we target a more challenging and realistic setting: open-set learning (OSL), where there exist test samples from the classes that are unseen during training. Although researchers have designed many methods from the algorithmic perspectives, there are few methods that provide generalization guarantees on their ability to achieve consistent performance on different training samples drawn from the same distribution. Motivated by the transfer learning and probably approximate correct (PAC) theory, we make a bold attempt to study OSL by proving its generalization error−given training samples with size n, the estimation error will get close to order Op(1/√n). This is the first study to provide a generalization bound for OSL, which we do by theoretically investigating the risk of the target classifier on unknown classes. According to our theory, a novel algorithm, called auxiliary open-set risk (AOSR) is proposed to address the OSL problem. Experiments verify the efficacy of AOSR. The code is available at github.com/AnjinLiu/OpensetLearningAOSR.

Learning to classify time series with limited data is a practical yet challenging problem. Current methods are primarily based on hand-designed feature extraction rules or domain-specific data augmentation. Motivated by the advances in deep speech processing models and the fact that voice data are univariate temporal signals, in this paper we propose Voice2Serie (V2S), a novel end-to-end approach that reprograms acoustic models for time series classification, through input transformation learning and output label mapping. Leveraging the representation learning power of a large-scale pre-trained speech processing model, on 31 different time series tasks we show that V2S outperforms or is on part with state-of-the-art methods on 22 tasks, and improves their average accuracy by 1.72%. We further provide theoretical justification of V2S by proving its population risk is upper bounded by the source risk and a Wasserstein distance accounting for feature alignment via reprogramming. Our results offer new and effective means to time series classification.

The current best practice for computing optimal transport (OT) is via entropy regularization and Sinkhorn iterations. This algorithm runs in quadratic time as it requires the full pairwise cost matrix, which is prohibitively expensive for large sets of objects. In this work we propose two effective log-linear time approximations of the cost matrix: First, a sparse approximation based on locality sensitive hashing (LSH) and, second, a Nyström approximation with LSH-based sparse corrections, which we call locally corrected Nyström (LCN). These approximations enable general log-linear time algorithms for entropy-regularized OT that perform well even for the complex, high-dimensional spaces common in deep learning. We analyse these approximations theoretically and evaluate them experimentally both directly and end-to-end as a component for real-world applications. Using our approximations for unsupervised word embedding alignment enables us to speed up a state-of-the-art method by a factor of 3 while also improving the accuracy by 3.1 percentage points without any additional model changes. For graph distance regression we propose the graph transport network (GTN), which combines graph neural networks (GNNs) with enhanced Sinkhorn. GTN outcompetes previous models by 48% and still scales log-linearly in the number of nodes.

The goal of metric learning is to learn a function that maps samples to a lower-dimensional space where similar samples lie closer than dissimilar ones. Particularly, deep metric learning utilizes neural networks to learn such a mapping. Most approaches rely on losses that only take the relations between pairs or triplets of samples into account, which either belong to the same class or two different classes. However, these methods do not explore the embedding space in its entirety. To this end, we propose an approach based on message passing networks that takes all the relations in a mini-batch into account. We refine embedding vectors by exchanging messages among all samples in a given batch allowing the training process to be aware of its overall structure. Since not all samples are equally important to predict a decision boundary, we use an attention mechanism during message passing to allow samples to weigh the importance of each neighbor accordingly. We achieve state-of-the-art results on clustering and image retrieval on the CUB-200-2011, Cars196, Stanford Online Products, and In-Shop Clothes datasets. To facilitate further research, we make available the code and the models at https://github.com/dvl-tum/intrabatchconnections.

We propose a general and scalable approximate sampling strategy for probabilistic models with discrete variables. Our approach uses gradients of the likelihood function with respect to its discrete inputs to propose updates in a Metropolis-Hastings sampler. We show empirically that this approach outperforms generic samplers in a number of difficult settings including Ising models, Potts models, restricted Boltzmann machines, and factorial hidden Markov models. We also demonstrate our improved sampler for training deep energy-based models on high dimensional discrete image data. This approach outperforms variational auto-encoders and existing energy-based models. Finally, we give bounds showing that our approach is near-optimal in the class of samplers which propose local updates.

Graph-structured data arise in a variety of real-world context ranging from sensor and transportation to biological and social networks. As a ubiquitous tool to process graph-structured data, spectral graph filters have been used to solve common tasks such as denoising and anomaly detection, as well as design deep learning architectures such as graph neural networks. Despite being an important tool, there is a lack of theoretical understanding of the stability properties of spectral graph filters, which are important for designing robust machine learning models. In this paper, we study filter stability and provide a novel and interpretable upper bound on the change of filter output, where the bound is expressed in terms of the endpoint degrees of the deleted and newly added edges, as well as the spatial proximity of those edges. This upper bound allows us to reason, in terms of structural properties of the graph, when a spectral graph filter will be stable. We further perform extensive experiments to verify intuition that can be gained from the bound.

Two-player, constant-sum games are well studied in the literature, but there has been limited progress outside of this setting. We propose Joint Policy-Space Response Oracles (JPSRO), an algorithm for training agents in n-player, general-sum extensive form games, which provably converges to an equilibrium. We further suggest correlated equilibria (CE) as promising meta-solvers, and propose a novel solution concept Maximum Gini Correlated Equilibrium (MGCE), a principled and computationally efficient family of solutions for solving the correlated equilibrium selection problem. We conduct several experiments using CE meta-solvers for JPSRO and demonstrate convergence on n-player, general-sum games.

Networks with node covariates are commonplace: for example, people in a social network have interests, or product preferences, etc. If we know the covariates for some nodes, can we infer them for the remaining nodes? In this paper we propose a new similarity measure between two nodes based on the patterns of their 2-hop neighborhoods. We show that a simple algorithm (CN-VEC) like nearest neighbor regression with this metric is consistent for a wide range of models when the degree grows faster than $n^{1/3}$ up-to logarithmic factors, where $n$ is the number of nodes. For "low-rank" latent variable models, the natural contender will be to estimate the latent variables using SVD and use them for non-parametric regression. While we show consistency of this method under less stringent sparsity conditions, our experimental results suggest that the simple local CN-VEC method either outperforms the global SVD-RBF method, or has comparable performance for low rank models. We also present simulated and real data experiments to show the effectiveness of our algorithms compared to the state of the art.

A novel asynchronous decentralized optimization method that follows Stochastic Variance Reduction (SVR) is proposed. Average consensus algorithms, such as Decentralized Stochastic Gradient Descent (DSGD), facilitate distributed training of machine learning models. However, the gradient will drift within the local nodes due to statistical heterogeneity of the subsets of data residing on the nodes and long communication intervals. To overcome the drift problem, (i) Gradient Tracking-SVR (GT-SVR) integrates SVR into DSGD and (ii) Edge-Consensus Learning (ECL) solves a model constrained minimization problem using a primal-dual formalism. In this paper, we reformulate the update procedure of ECL such that it implicitly includes the gradient modification of SVR by optimally selecting a constraint-strength control parameter. Through convergence analysis and experiments, we confirmed that the proposed ECL with Implicit SVR (ECL-ISVR) is stable and approximately reaches the reference performance obtained with computation on a single-node using full data set.

Successful negotiators must learn how to balance optimizing for self-interest and cooperation. Yet current artificial negotiation agents often heavily depend on the quality of the static datasets they were trained on, limiting their capacity to fashion an adaptive response balancing self-interest and cooperation. For this reason, we find that these agents can achieve either high utility or cooperation, but not both. To address this, we introduce a targeted data acquisition framework where we guide the exploration of a reinforcement learning agent using annotations from an expert oracle. The guided exploration incentivizes the learning agent to go beyond its static dataset and develop new negotiation strategies. We show that this enables our agents to obtain higher-reward and more Pareto-optimal solutions when negotiating with both simulated and human partners compared to standard supervised learning and reinforcement learning methods. This trend additionally holds when comparing agents using our targeted data acquisition framework to variants of agents trained with a mix of supervised learning and reinforcement learning, or to agents using tailored reward functions that explicitly optimize for utility and Pareto-optimality.

In this paper we consider the problem of learning an $\epsilon$-optimal policy for a discounted Markov Decision Process (MDP). Given an MDP with $S$ states, $A$ actions, the discount factor $\gamma \in (0,1)$, and an approximation threshold $\epsilon > 0$, we provide a model-free algorithm to learn an $\epsilon$-optimal policy with sample complexity $\tilde{O}(\frac{SA\ln(1/p)}{\epsilon^2(1-\gamma)^{5.5}})$ \footnote{In this work, the notation $\tilde{O}(\cdot)$ hides poly-logarithmic factors of $S,A,1/(1-\gamma)$, and $1/\epsilon$.} and success probability $(1-p)$. For small enough $\epsilon$, we show an improved algorithm with sample complexity $\tilde{O}(\frac{SA\ln(1/p)}{\epsilon^2(1-\gamma)^{3}})$. While the first bound improves upon all known model-free algorithms and model-based ones with tight dependence on $S$, our second algorithm beats all known sample complexity bounds and matches the information theoretic lower bound up to logarithmic factors.

The training of deep residual neural networks (ResNets) with backpropagation has a memory cost that increases linearly with respect to the depth of the network. A simple way to circumvent this issue is to use reversible architectures. In this paper, we propose to change the forward rule of a ResNet by adding a momentum term. The resulting networks, momentum residual neural networks (MomentumNets), are invertible. Unlike previous invertible architectures, they can be used as a drop-in replacement for any existing ResNet block. We show that MomentumNets can be interpreted in the infinitesimal step size regime as second-order ordinary differential equations (ODEs) and exactly characterize how adding momentum progressively increases the representation capabilities of MomentumNets: they can learn any linear mapping up to a multiplicative factor, while ResNets cannot. In a learning to optimize setting, where convergence to a fixed point is required, we show theoretically and empirically that our method succeeds while existing invertible architectures fail. We show on CIFAR and ImageNet that MomentumNets have the same accuracy as ResNets, while having a much smaller memory footprint, and show that pre-trained MomentumNets are promising for fine-tuning models.

Current model-based reinforcement learning methods struggle when operating from complex visual scenes due to their inability to prioritize task-relevant features. To mitigate this problem, we propose learning Task Informed Abstractions (TIA) that explicitly separates reward-correlated visual features from distractors. For learning TIA, we introduce the formalism of Task Informed MDP (TiMDP) that is realized by training two models that learn visual features via cooperative reconstruction, but one model is adversarially dissociated from the reward signal. Empirical evaluation shows that TIA leads to significant performance gains over state-of-the-art methods on many visual control tasks where natural and unconstrained visual distractions pose a formidable challenge. Project page: https://xiangfu.co/tia

Symmetric functions, which take as input an unordered, fixed-size set, are known to be universally representable by neural networks that enforce permutation invariance. These architectures only give guarantees for fixed input sizes, yet in many practical applications, including point clouds and particle physics, a relevant notion of generalization should include varying the input size. In this work we treat symmetric functions (of any size) as functions over probability measures, and study the learning and representation of neural networks defined on measures. By focusing on shallow architectures, we establish approximation and generalization bounds under different choices of regularization (such as RKHS and variation norms), that capture a hierarchy of functional spaces with increasing degree of non-linear learning. The resulting models can be learned efficiently and enjoy generalization guarantees that extend across input sizes, as we verify empirically.

We introduce the GANsformer, a novel and efficient type of transformer, and explore it for the task of visual generative modeling. The network employs a bipartite structure that enables long-range interactions across the image, while maintaining computation of linear efficiency, that can readily scale to high-resolution synthesis. It iteratively propagates information from a set of latent variables to the evolving visual features and vice versa, to support the refinement of each in light of the other and encourage the emergence of compositional representations of objects and scenes. In contrast to the classic transformer architecture, it utilizes multiplicative integration that allows flexible region-based modulation, and can thus be seen as a generalization of the successful StyleGAN network. We demonstrate the model's strength and robustness through a careful evaluation over a range of datasets, from simulated multi-object environments to rich real-world indoor and outdoor scenes, showing it achieves state-of-the-art results in terms of image quality and diversity, while enjoying fast learning and better data-efficiency. Further qualitative and quantitative experiments offer us an insight into the model's inner workings, revealing improved interpretability and stronger disentanglement, and illustrating the benefits and efficacy of our approach. An implementation of the model is available at https://github.com/dorarad/gansformer.

The Bayesian paradigm has the potential to solve core issues of deep neural networks such as poor calibration and data inefficiency. Alas, scaling Bayesian inference to large weight spaces often requires restrictive approximations. In this work, we show that it suffices to perform inference over a small subset of model weights in order to obtain accurate predictive posteriors. The other weights are kept as point estimates. This subnetwork inference framework enables us to use expressive, otherwise intractable, posterior approximations over such subsets. In particular, we implement subnetwork linearized Laplace as a simple, scalable Bayesian deep learning method: We first obtain a MAP estimate of all weights and then infer a full-covariance Gaussian posterior over a subnetwork using the linearized Laplace approximation. We propose a subnetwork selection strategy that aims to maximally preserve the model’s predictive uncertainty. Empirically, our approach compares favorably to ensembles and less expressive posterior approximations over full networks.

What is the computational model behind a Transformer? Where recurrent neural networks have direct parallels in finite state machines, allowing clear discussion and thought around architecture variants or trained models, Transformers have no such familiar parallel. In this paper we aim to change that, proposing a computational model for the transformer-encoder in the form of a programming language. We map the basic components of a transformer-encoder---attention and feed-forward computation---into simple primitives, around which we form a programming language: the Restricted Access Sequence Processing Language (RASP). We show how RASP can be used to program solutions to tasks that could conceivably be learned by a Transformer, and how a Transformer can be trained to mimic a RASP solution. In particular, we provide RASP programs for histograms, sorting, and Dyck-languages. We further use our model to relate their difficulty in terms of the number of required layers and attention heads: analyzing a RASP program implies a maximum number of heads and layers necessary to encode a task in a transformer. Finally, we see how insights gained from our abstraction might be used to explain phenomena seen in recent works.

We present a method for efficient differentiable simulation of articulated bodies. This enables integration of articulated body dynamics into deep learning frameworks, and gradient-based optimization of neural networks that operate on articulated bodies. We derive the gradients of the contact solver using spatial algebra and the adjoint method. Our approach is an order of magnitude faster than autodiff tools. By only saving the initial states throughout the simulation process, our method reduces memory requirements by two orders of magnitude. We demonstrate the utility of efficient differentiable dynamics for articulated bodies in a variety of applications. We show that reinforcement learning with articulated systems can be accelerated using gradients provided by our method. In applications to control and inverse problems, gradient-based optimization enabled by our work accelerates convergence by more than an order of magnitude.

The design of recurrent neural networks (RNNs) to accurately process sequential inputs with long-time dependencies is very challenging on account of the exploding and vanishing gradient problem. To overcome this, we propose a novel RNN architecture which is based on a structure preserving discretization of a Hamiltonian system of second-order ordinary differential equations that models networks of oscillators. The resulting RNN is fast, invertible (in time), memory efficient and we derive rigorous bounds on the hidden state gradients to prove the mitigation of the exploding and vanishing gradient problem. A suite of experiments are presented to demonstrate that the proposed RNN provides state of the art performance on a variety of learning tasks with (very) long-time dependencies.

Deep learning has made revolutionary advances to diverse applications in the presence of large-scale labeled datasets. However, it is prohibitively time-costly and labor-expensive to collect sufficient labeled data in most realistic scenarios. To mitigate the requirement for labeled data, semi-supervised learning (SSL) focuses on simultaneously exploring both labeled and unlabeled data, while transfer learning (TL) popularizes a favorable practice of fine-tuning a pre-trained model to the target data. A dilemma is thus encountered: Without a decent pre-trained model to provide an implicit regularization, SSL through self-training from scratch will be easily misled by inaccurate pseudo-labels, especially in large-sized label space; Without exploring the intrinsic structure of unlabeled data, TL through fine-tuning from limited labeled data is at risk of under-transfer caused by model shift. To escape from this dilemma, we present Self-Tuning to enable data-efficient deep learning by unifying the exploration of labeled and unlabeled data and the transfer of a pre-trained model, as well as a Pseudo Group Contrast (PGC) mechanism to mitigate the reliance on pseudo-labels and boost the tolerance to false labels. Self-Tuning outperforms its SSL and TL counterparts on five tasks by sharp margins, e.g. it doubles the accuracy of fine-tuning on Cars with $15\%$ labels.

In contrast to traditional weight optimization in a continuous space, we demonstrate the existence of effective random networks whose weights are never updated. By selecting a weight among a fixed set of random values for each individual connection, our method uncovers combinations of random weights that match the performance of traditionally-trained networks of the same capacity. We refer to our networks as "slot machines" where each reel (connection) contains a fixed set of symbols (random values). Our backpropagation algorithm "spins" the reels to seek "winning" combinations, i.e., selections of random weight values that minimize the given loss. Quite surprisingly, we find that allocating just a few random values to each connection (e.g., 8 values per connection) yields highly competitive combinations despite being dramatically more constrained compared to traditionally learned weights. Moreover, finetuning these combinations often improves performance over the trained baselines. A randomly initialized VGG-19 with 8 values per connection contains a combination that achieves 91% test accuracy on CIFAR-10. Our method also achieves an impressive performance of 98.2% on MNIST for neural networks containing only random weights.

We present a novel interactive learning protocol that enables training request-fulfilling agents by verbally describing their activities. Unlike imitation learning (IL), our protocol allows the teaching agent to provide feedback in a language that is most appropriate for them. Compared with reward in reinforcement learning (RL), the description feedback is richer and allows for improved sample complexity. We develop a probabilistic framework and an algorithm that practically implements our protocol. Empirical results in two challenging request-fulfilling problems demonstrate the strengths of our approach: compared with RL baselines, it is more sample-efficient; compared with IL baselines, it achieves competitive success rates without requiring the teaching agent to be able to demonstrate the desired behavior using the learning agent’s actions. Apart from empirical evaluation, we also provide theoretical guarantees for our algorithm under certain assumptions about the teacher and the environment.

Standard deep reinforcement learning algorithms use a shared representation for the policy and value function, especially when training directly from images. However, we argue that more information is needed to accurately estimate the value function than to learn the optimal policy. Consequently, the use of a shared representation for the policy and value function can lead to overfitting. To alleviate this problem, we propose two approaches which are combined to create IDAAC: Invariant Decoupled Advantage Actor-Critic. First, IDAAC decouples the optimization of the policy and value function, using separate networks to model them. Second, it introduces an auxiliary loss which encourages the representation to be invariant to task-irrelevant properties of the environment. IDAAC shows good generalization to unseen environments, achieving a new state-of-the-art on the Procgen benchmark and outperforming popular methods on DeepMind Control tasks with distractors. Our implementation is available at https://github.com/rraileanu/idaac.

We consider a distributed non-convex optimization where a network of agents aims at minimizing a global function over the Stiefel manifold. The global function is represented as a finite sum of smooth local functions, where each local function is associated with one agent and agents communicate with each other over an undirected connected graph. The problem is non-convex as local functions are possibly non-convex (but smooth) and the Steifel manifold is a non-convex set. We present a decentralized Riemannian stochastic gradient method (DRSGD) with the convergence rate of $\mathcal{O}(1/\sqrt{K})$ to a stationary point. To have exact convergence with constant stepsize, we also propose a decentralized Riemannian gradient tracking algorithm (DRGTA) with the convergence rate of $\mathcal{O}(1/K)$ to a stationary point. We use multi-step consensus to preserve the iteration in the local (consensus) region. DRGTA is the first decentralized algorithm with exact convergence for distributed optimization on Stiefel manifold.

We investigate the distributions of Conjugate Kernel (CK) and Neural Tangent Kernel (NTK) for ReLU networks with random initialization. We derive the precise distributions and moments of the diagonal elements of these kernels. For a feedforward network, these values converge in law to a log-normal distribution when the network depth $d$ and width $n$ simultaneously tend to infinity and the variance of log diagonal elements is proportional to ${d}/{n}$. For the residual network, in the limit that number of branches $m$ increases to infinity and the width $n$ remains fixed, the diagonal elements of Conjugate Kernel converge in law to a log-normal distribution where the variance of log value is proportional to ${1}/{n}$, and the diagonal elements of NTK converge in law to a log-normal distributed variable times the conjugate kernel of one feedforward network. Our new theoretical analysis results suggest that residual network remains trainable in the limit of infinite branches and fixed network width. The numerical experiments are conducted and all results validate the soundness of our theoretical analysis.

The convergence rate of Value Iteration (VI), a fundamental procedure in dynamic programming and reinforcement learning, for solving MDPs can be slow when the discount factor is close to one. We propose modifications to VI in order to potentially accelerate its convergence behaviour. The key insight is the realization that the evolution of the value function approximations $(V_k)_{k \geq 0}$ in the VI procedure can be seen as a dynamical system. This opens up the possibility of using techniques from \emph{control theory} to modify, and potentially accelerate, this dynamics. We present such modifications based on simple controllers, such as PD (Proportional-Derivative), PI (Proportional-Integral), and PID. We present the error dynamics of these variants of VI, and provably (for certain classes of MDPs) and empirically (for more general classes) show that the convergence rate can be significantly improved. We also propose a gain adaptation mechanism in order to automatically select the controller gains, and empirically show the effectiveness of this procedure.

We introduce a new neural network architecture that we call "grid-functioned" neural networks. It utilises a grid structure of network parameterisations that can be specialised for different subdomains of the problem, while maintaining smooth, continuous behaviour. The grid gives the user flexibility to prevent gross features from overshadowing important minor ones. We present a full characterisation of its computational and spatial complexity, and demonstrate its potential, compared to a traditional architecture, over a set of synthetic regression problems. We further illustrate the benefits through a real-world 3D skeletal animation case study, where it offers the same visual quality as a state-of-the-art model, but with lower computational complexity and better control accuracy.

We address the problem of convex optimization with preference (dueling) feedback. Like the traditional optimization objective, the goal is to find the optimal point with the least possible query complexity, however, without the luxury of even a zeroth order feedback. Instead, the learner can only observe a single noisy bit which is win-loss feedback for a pair of queried points based on their function values. % The problem is certainly of great practical relevance as in many real-world scenarios, such as recommender systems or learning from customer preferences, where the system feedback is often restricted to just one binary-bit preference information. % We consider the problem of online convex optimization (OCO) solely by actively querying $\{0,1\}$ noisy-comparison feedback of decision point pairs, with the objective of finding a near-optimal point (function minimizer) with the least possible number of queries. %a very general class of monotonic, non-decreasing transfer functions, and analyze the problem for any $d$-dimensional smooth convex function. % For the non-stationary OCO setup, where the underlying convex function may change over time, we prove an impossibility result towards achieving the above objective. We next focus only on the stationary OCO problem, and our main contribution lies in designing a normalized gradient descent based algorithm towards finding a $\epsilon$-best optimal point. Towards this, our algorithm is shown to yield a convergence rate of $\tilde O(\nicefrac{d\beta}{\epsilon \nu^2})$ ($\nu$ being the noise parameter) when the underlying function is $\beta$-smooth. Further we show an improved convergence rate of just $\tilde O(\nicefrac{d\beta}{\alpha \nu^2} \log \frac{1}{\epsilon})$ when the function is additionally also $\alpha$-strongly convex.

Although stochastic optimization is central to modern machine learning, the precise mechanisms underlying its success, and in particular, the precise role of the stochasticity, still remain unclear. Modeling stochastic optimization algorithms as discrete random recurrence relations, we show that multiplicative noise, as it commonly arises due to variance in local rates of convergence, results in heavy-tailed stationary behaviour in the parameters. Theoretical results are obtained characterizing this for a large class of (non-linear and even non-convex) models and optimizers (including momentum, Adam, and stochastic Newton), demonstrating that this phenomenon holds generally. We describe dependence on key factors, including step size, batch size, and data variability, all of which exhibit similar qualitative behavior to recent empirical results on state-of-the-art neural network models. Furthermore, we empirically illustrate how multiplicative noise and heavy-tailed structure improve capacity for basin hopping and exploration of non-convex loss surfaces, over commonly-considered stochastic dynamics with only additive noise and light-tailed structure.

While deep neural networks provide good performance for a range of challenging tasks, calibration and uncertainty estimation remain major challenges, especially under distribution shift. In this paper, we propose the amortized conditional normalized maximum likelihood (ACNML) method as a scalable general-purpose approach for uncertainty estimation, calibration, and out-of-distribution robustness with deep networks. Our algorithm builds on the conditional normalized maximum likelihood (CNML) coding scheme, which has minimax optimal properties according to the minimum description length principle, but is computationally intractable to evaluate exactly for all but the simplest of model classes. We propose to use approximate Bayesian inference technqiues to produce a tractable approximation to the CNML distribution. Our approach can be combined with any approximate inference algorithm that provides tractable posterior densities over model parameters. We demonstrate that ACNML compares favorably to a number of prior techniques for uncertainty estimation in terms of calibration when faced with distribution shift.

Attention mechanisms have shown promising results in sequence modeling tasks that require long-term memory. Recent work investigated mechanisms to reduce the computational cost of preserving and storing memories. However, not all content in the past is equally important to remember. We propose Expire-Span, a method that learns to retain the most important information and expire the irrelevant information. This forgetting of memories enables Transformers to scale to attend over tens of thousands of previous timesteps efficiently, as not all states from previous timesteps are preserved. We demonstrate that Expire-Span can help models identify and retain critical information and show it can achieve strong performance on reinforcement learning tasks specifically designed to challenge this functionality. Next, we show that Expire-Span can scale to memories that are tens of thousands in size, setting a new state of the art on incredibly long context tasks such as character-level language modeling and a frame-by-frame moving objects task. Finally, we analyze the efficiency of Expire-Span compared to existing approaches and demonstrate that it trains faster and uses less memory.

Selective labels are a common feature of high-stakes decision-making applications, referring to the lack of observed outcomes under one of the possible decisions. This paper studies the learning of decision policies in the face of selective labels, in an online setting that balances learning costs against future utility. In the homogeneous case in which individuals' features are disregarded, the optimal decision policy is shown to be a threshold policy. The threshold becomes more stringent as more labels are collected; the rate at which this occurs is characterized. In the case of features drawn from a finite domain, the optimal policy consists of multiple homogeneous policies in parallel. For the general infinite-domain case, the homogeneous policy is extended by using a probabilistic classifier and bootstrapping to provide its inputs. In experiments on synthetic and real data, the proposed policies achieve consistently superior utility with no parameter tuning in the finite-domain case and lower parameter sensitivity in the general case.

In real world settings, numerous constraints are present which are hard to specify mathematically. However, for the real world deployment of reinforcement learning (RL), it is critical that RL agents are aware of these constraints, so that they can act safely. In this work, we consider the problem of learning constraints from demonstrations of a constraint-abiding agent's behavior. We experimentally validate our approach and show that our framework can successfully learn the most likely constraints that the agent respects. We further show that these learned constraints are \textit{transferable} to new agents that may have different morphologies and/or reward functions. Previous works in this regard have either mainly been restricted to tabular (discrete) settings, specific types of constraints or assume the environment's transition dynamics. In contrast, our framework is able to learn arbitrary \textit{Markovian} constraints in high-dimensions in a completely model-free setting. The code is available at: \url{https://github.com/shehryar-malik/icrl}.

A graph generative model defines a distribution over graphs. Typically, the model consists of a sequential process that creates and adds nodes and edges. Such sequential process defines an ordering of the nodes in the graph. The computation of the model's likelihood requires to marginalize the node orderings; this makes maximum likelihood estimation (MLE) challenging due to the (factorial) number of possible permutations. In this work, we provide an expression for the likelihood of a graph generative model and show that its calculation is closely related to the problem of graph automorphism. In addition, we derive a variational inference (VI) algorithm for fitting a graph generative model that is based on the maximization of a variational bound of the log-likelihood. This allows the model to be trained with node orderings from the approximate posterior instead of ad-hoc orderings. Our experiments show that our log-likelihood bound is significantly tighter than the bound of previous schemes. The models fitted with the VI algorithm are able to generate high-quality graphs that match the structures of target graphs not seen during training.

Deep generative models have been demonstrated as state-of-the-art density estimators. Yet, recent work has found that they often assign a higher likelihood to data from outside the training distribution. This seemingly paradoxical behavior has caused concerns over the quality of the attained density estimates. In the context of hierarchical variational autoencoders, we provide evidence to explain this behavior by out-of-distribution data having in-distribution low-level features. We argue that this is both expected and desirable behavior. With this insight in hand, we develop a fast, scalable and fully unsupervised likelihood-ratio score for OOD detection that requires data to be in-distribution across all feature-levels. We benchmark the method on a vast set of data and model combinations and achieve state-of-the-art results on out-of-distribution detection.

We introduce the Graph Mixture Density Networks, a new family of machine learning models that can fit multimodal output distributions conditioned on graphs of arbitrary topology. By combining ideas from mixture models and graph representation learning, we address a broader class of challenging conditional density estimation problems that rely on structured data. In this respect, we evaluate our method on a new benchmark application that leverages random graphs for stochastic epidemic simulations. We show a significant improvement in the likelihood of epidemic outcomes when taking into account both multimodality and structure. The empirical analysis is complemented by two real-world regression tasks showing the effectiveness of our approach in modeling the output prediction uncertainty. Graph Mixture Density Networks open appealing research opportunities in the study of structure-dependent phenomena that exhibit non-trivial conditional output distributions.

Marginal-likelihood based model-selection, even though promising, is rarely used in deep learning due to estimation difficulties. Instead, most approaches rely on validation data, which may not be readily available. In this work, we present a scalable marginal-likelihood estimation method to select both hyperparameters and network architectures, based on the training data alone. Some hyperparameters can be estimated online during training, simplifying the procedure. Our marginal-likelihood estimate is based on Laplace’s method and Gauss-Newton approximations to the Hessian, and it outperforms cross-validation and manual tuning on standard regression and image classification datasets, especially in terms of calibration and out-of-distribution detection. Our work shows that marginal likelihoods can improve generalization and be useful when validation data is unavailable (e.g., in nonstationary settings).

Detecting the Maximum Common Subgraph (MCS) between two input graphs is fundamental for applications in drug synthesis, malware detection, cloud computing, etc. However, MCS computation is NP-hard, and state-of-the-art MCS solvers rely on heuristic search algorithms which in practice cannot find good solution for large graph pairs given a limited computation budget. We propose GLSearch, a Graph Neural Network (GNN) based learning to search model. Our model is built upon the branch and bound algorithm, which selects one pair of nodes from the two input graphs to expand at a time. We propose a novel GNN-based Deep Q-Network (DQN) to select the node pair, making the search process much faster. Experiments on synthetic and real-world graph pairs demonstrate that our model learns a search strategy that is able to detect significantly larger common subgraphs than existing MCS solvers given the same computation budget. GLSearch can be potentially extended to solve many other combinatorial problems with constraints on graphs.

Learning sensorimotor control policies from high-dimensional images crucially relies on the quality of the underlying visual representations. Prior works show that structured latent space such as visual keypoints often outperforms unstructured representations for robotic control. However, most of these representations, whether structured or unstructured are learned in a 2D space even though the control tasks are usually performed in a 3D environment. In this work, we propose a framework to learn such a 3D geometric structure directly from images in an end-to-end unsupervised manner. The input images are embedded into latent 3D keypoints via a differentiable encoder which is trained to optimize both a multi-view consistency loss and downstream task objective. These discovered 3D keypoints tend to meaningfully capture robot joints as well as object movements in a consistent manner across both time and 3D space. The proposed approach outperforms prior state-of-art methods across a variety of reinforcement learning benchmarks. Code and videos at https://buoyancy99.github.io/unsup-3d-keypoints/.

In few-shot imitation, an agent is given a few demonstrations of a previously unseen task, and must then successfully perform that task. We propose a novel approach to learning few-shot-imitation agents that we call demonstration-conditioned reinforcement learning (DCRL). Given a training set consisting of demonstrations, reward functions and transition distributions for multiple tasks, the idea is to work with a policy that takes demonstrations as input, and to train this policy to maximize the average of the cumulative reward over the set of training tasks. Relative to previously proposed few-shot imitation methods that use behaviour cloning or infer reward functions from demonstrations, our method has the disadvantage that it requires reward functions at training time. However, DCRL also has several advantages, such as the ability to improve upon suboptimal demonstrations, to operate given state-only demonstrations, and to cope with a domain shift between the demonstrator and the agent. Moreover, we show that DCRL outperforms methods based on behaviour cloning by a large margin, on navigation tasks and on robotic manipulation tasks from the Meta-World benchmark.

Learning effective representations in image-based environments is crucial for sample efficient Reinforcement Learning (RL). Unfortunately, in RL, representation learning is confounded with the exploratory experience of the agent -- learning a useful representation requires diverse data, while effective exploration is only possible with coherent representations. Furthermore, we would like to learn representations that not only generalize across tasks but also accelerate downstream exploration for efficient task-specific training. To address these challenges we propose Proto-RL, a self-supervised framework that ties representation learning with exploration through prototypical representations. These prototypes simultaneously serve as a summarization of the exploratory experience of an agent as well as a basis for representing observations. We pre-train these task-agnostic representations and prototypes on environments without downstream task information. This enables state-of-the-art downstream policy learning on a set of difficult continuous control tasks.

With the development of deep networks on various large-scale datasets, a large zoo of pretrained models are available. When transferring from a model zoo, applying classic single-model-based transfer learning methods to each source model suffers from high computational cost and cannot fully utilize the rich knowledge in the zoo. We propose \emph{Zoo-Tuning} to address these challenges, which learns to adaptively transfer the parameters of pretrained models to the target task. With the learnable channel alignment layer and adaptive aggregation layer, Zoo-Tuning \emph{adaptively aggregates channel aligned pretrained parameters to derive the target model}, which simultaneously promotes knowledge transfer and adapts source models to downstream tasks. The adaptive aggregation substantially reduces the computation cost at both training and inference. We further propose lite Zoo-Tuning with the temporal ensemble of batch average gating values to reduce the storage cost at the inference time. We evaluate our approach on a variety of tasks, including reinforcement learning, image classification, and facial landmark detection. Experiment results demonstrate that the proposed adaptive transfer learning approach can more effectively and efficiently transfer knowledge from a zoo of models.

Since the Message Passing (Graph) Neural Networks (MPNNs) have a linear complexity with respect to the number of nodes when applied to sparse graphs, they have been widely implemented and still raise a lot of interest even though their theoretical expressive power is limited to the first order Weisfeiler-Lehman test (1-WL). In this paper, we show that if the graph convolution supports are designed in spectral-domain by a non-linear custom function of eigenvalues and masked with an arbitrary large receptive field, the MPNN is theoretically more powerful than the 1-WL test and experimentally as powerful as a 3-WL existing models, while remaining spatially localized. Moreover, by designing custom filter functions, outputs can have various frequency components that allow the convolution process to learn different relationships between a given input graph signal and its associated properties. So far, the best 3-WL equivalent graph neural networks have a computational complexity in $\mathcal{O}(n^3)$ with memory usage in $\mathcal{O}(n^2)$, consider non-local update mechanism and do not provide the spectral richness of output profile. The proposed method overcomes all these aforementioned problems and reaches state-of-the-art results in many downstream tasks.

Understanding the fundamental mechanism behind the success of deep neural networks is one of the key challenges in the modern machine learning literature. Despite numerous attempts, a solid theoretical analysis is yet to be developed. In this paper, we develop a novel unified framework to reveal a hidden regularization mechanism through the lens of convex optimization. We first show that the training of multiple three-layer ReLU sub-networks with weight decay regularization can be equivalently cast as a convex optimization problem in a higher dimensional space, where sparsity is enforced via a group $\ell_1$-norm regularization. Consequently, ReLU networks can be interpreted as high dimensional feature selection methods. More importantly, we then prove that the equivalent convex problem can be globally optimized by a standard convex optimization solver with a polynomial-time complexity with respect to the number of samples and data dimension when the width of the network is fixed. Finally, we numerically validate our theoretical results via experiments involving both synthetic and real datasets.

We propose a generative model that can infer a distribution for the underlying spatial signal conditioned on sparse samples e.g. plausible images given a few observed pixels. In contrast to sequential autoregressive generative models, our model allows conditioning on arbitrary samples and can answer distributional queries for any location. We empirically validate our approach across three image datasets and show that we learn to generate diverse and meaningful samples, with the distribution variance reducing given more observed pixels. We also show that our approach is applicable beyond images and can allow generating other types of spatial outputs e.g. polynomials, 3D shapes, and videos.

We propose a novel policy update that combines regularized policy optimization with model learning as an auxiliary loss. The update (henceforth Muesli) matches MuZero's state-of-the-art performance on Atari. Notably, Muesli does so without using deep search: it acts directly with a policy network and has computation speed comparable to model-free baselines. The Atari results are complemented by extensive ablations, and by additional results on continuous control and 9x9 Go.

It is well-known that stochastic gradient noise (SGN) acts as implicit regularization for deep learning and is essentially important for both optimization and generalization of deep networks. Some works attempted to artificially simulate SGN by injecting random noise to improve deep learning. However, it turned out that the injected simple random noise cannot work as well as SGN, which is anisotropic and parameter-dependent. For simulating SGN at low computational costs and without changing the learning rate or batch size, we propose the Positive-Negative Momentum (PNM) approach that is a powerful alternative to conventional Momentum in classic optimizers. The introduced PNM method maintains two approximate independent momentum terms. Then, we can control the magnitude of SGN explicitly by adjusting the momentum difference. We theoretically prove the convergence guarantee and the generalization advantage of PNM over Stochastic Gradient Descent (SGD). By incorporating PNM into the two conventional optimizers, SGD with Momentum and Adam, our extensive experiments empirically verified the significant advantage of the PNM-based variants over the corresponding conventional Momentum-based optimizers. Code: \url{https://github.com/zeke-xie/Positive-Negative-Momentum}.

In this paper, sample-aware policy entropy regularization is proposed to enhance the conventional policy entropy regularization for better exploration. Exploiting the sample distribution obtainable from the replay buffer, the proposed sample-aware entropy regularization maximizes the entropy of the weighted sum of the policy action distribution and the sample action distribution from the replay buffer for sample-efficient exploration. A practical algorithm named diversity actor-critic (DAC) is developed by applying policy iteration to the objective function with the proposed sample-aware entropy regularization. Numerical results show that DAC significantly outperforms existing recent algorithms for reinforcement learning.

It is widely believed that engineering a model to be invariant/equivariant improves generalisation. Despite the growing popularity of this approach, a precise characterisation of the generalisation benefit is lacking. By considering the simplest case of linear models, this paper provides the first provably non-zero improvement in generalisation for invariant/equivariant models when the target distribution is invariant/equivariant with respect to a compact group. Moreover, our work reveals an interesting relationship between generalisation, the number of training examples and properties of the group action. Our results rest on an observation of the structure of function spaces under averaging operators which, along with its consequences for feature averaging, may be of independent interest.

Optimal transport distances have found many applications in machine learning for their capacity to compare non-parametric probability distributions. Yet their algorithmic complexity generally prevents their direct use on large scale datasets. Among the possible strategies to alleviate this issue, practitioners can rely on computing estimates of these distances over subsets of data, i.e. minibatches. While computationally appealing, we highlight in this paper some limits of this strategy, arguing it can lead to undesirable smoothing effects. As an alternative, we suggest that the same minibatch strategy coupled with unbalanced optimal transport can yield more robust behaviors. We discuss the associated theoretical properties, such as unbiased estimators, existence of gradients and concentration bounds. Our experimental study shows that in challenging problems associated to domain adaptation, the use of unbalanced optimal transport leads to significantly better results, competing with or surpassing recent baselines.

In general, graph representation learning methods assume that the train and test data come from the same distribution. In this work we consider an underexplored area of an otherwise rapidly developing field of graph representation learning: The task of out-of-distribution (OOD) graph classification, where train and test data have different distributions, with test data unavailable during training. Our work shows it is possible to use a causal model to learn approximately invariant representations that better extrapolate between train and test data. Finally, we conclude with synthetic and real-world dataset experiments showcasing the benefits of representations that are invariant to train/test distribution shifts.

Recent exploration methods have proven to be a recipe for improving sample-efficiency in deep reinforcement learning (RL). However, efficient exploration in high-dimensional observation spaces still remains a challenge. This paper presents Random Encoders for Efficient Exploration (RE3), an exploration method that utilizes state entropy as an intrinsic reward. In order to estimate state entropy in environments with high-dimensional observations, we utilize a k-nearest neighbor entropy estimator in the low-dimensional representation space of a convolutional encoder. In particular, we find that the state entropy can be estimated in a stable and compute-efficient manner by utilizing a randomly initialized encoder, which is fixed throughout training. Our experiments show that RE3 significantly improves the sample-efficiency of both model-free and model-based RL methods on locomotion and navigation tasks from DeepMind Control Suite and MiniGrid benchmarks. We also show that RE3 allows learning diverse behaviors without extrinsic rewards, effectively improving sample-efficiency in downstream tasks.

The Dirichlet Belief Network~(DirBN) was recently proposed as a promising deep generative model to learn interpretable deep latent distributions for objects. However, its current representation capability is limited since its latent distributions across different layers is prone to form similar patterns and can thus hardly use multi-layer structure to form flexible distributions. In this work, we propose Poisson-randomised Dirichlet Belief Networks (Pois-DirBN), which allows large mutations for the latent distributions across layers to enlarge the representation capability. Based on our key idea of inserting Poisson random variables in the layer-wise connection, Pois-DirBN first introduces a component-wise propagation mechanism to enable latent distributions to have large variations across different layers. Then, we develop a layer-wise Gibbs sampling algorithm to infer the latent distributions, leading to a larger number of effective layers compared to DirBN. In addition, we integrate out latent distributions and form a multi-stochastic deep integer network, which provides an alternative view on Pois-DirBN. We apply Pois-DirBN to relational modelling and validate its effectiveness through improved link prediction performance and more interpretable latent distribution visualisations. The code can be downloaded at https://github.com/xuhuifan/Pois_DirBN.

In real-world multi-agent systems, agents with different capabilities may join or leave without altering the team's overarching goals. Coordinating teams with such dynamic composition is challenging: the optimal team strategy varies with the composition. We propose COPA, a coach-player framework to tackle this problem. We assume the coach has a global view of the environment and coordinates the players, who only have partial views, by distributing individual strategies. Specifically, we 1) adopt the attention mechanism for both the coach and the players; 2) propose a variational objective to regularize learning; and 3) design an adaptive communication method to let the coach decide when to communicate with the players. We validate our methods on a resource collection task, a rescue game, and the StarCraft micromanagement tasks. We demonstrate zero-shot generalization to new team compositions. Our method achieves comparable or better performance than the setting where all players have a full view of the environment. Moreover, we see that the performance remains high even when the coach communicates as little as 13% of the time using the adaptive communication strategy.

Sparsifying deep neural networks is of paramount interest in many areas, especially when those networks have to be implemented on low-memory devices. In this article, we propose a new formulation of the problem of generating sparse weights for a pre-trained neural network. By leveraging the properties of standard nonlinear activation functions, we show that the problem is equivalent to an approximate subdifferential inclusion problem. The accuracy of the approximation controls the sparsity. We show that the proposed approach is valid for a broad class of activation functions (ReLU, sigmoid, softmax). We propose an iterative optimization algorithm to induce sparsity whose convergence is guaranteed. Because of the algorithm flexibility, the sparsity can be ensured from partial training data in a minibatch manner. To demonstrate the effectiveness of our method, we perform experiments on various networks in different applicative contexts: image classification, speech recognition, natural language processing, and time-series forecasting.

The reward function is widely accepted as a succinct, robust, and transferable representation of a task. Typical approaches, at the basis of Inverse Reinforcement Learning (IRL), leverage on expert demonstrations to recover a reward function. In this paper, we study the theoretical properties of the class of reward functions that are compatible with the expert’s behavior. We analyze how the limited knowledge of the expert’s policy and of the environment affects the reward reconstruction phase. Then, we examine how the error propagates to the learned policy’s performance when transferring the reward function to a different environment. We employ these findings to devise a provably efficient active sampling approach, aware of the need for transferring the reward function, that can be paired with a large variety of IRL algorithms. Finally, we provide numerical simulations on benchmark environments.

This paper considers classification problems with hierarchically organized classes. We force the classifier (hyperplane) of each class to belong to a sphere manifold, whose center is the classifier of its super-class. Then, individual sphere manifolds are connected based on their hierarchical relations. Our technique replaces the last layer of a neural network by combining a spherical fully-connected layer with a hierarchical layer. This regularization is shown to improve the performance of widely used deep neural network architectures (ResNet and DenseNet) on publicly available datasets (CIFAR100, CUB200, Stanford dogs, Stanford cars, and Tiny-ImageNet).

We introduce a new skill-discovery algorithm that builds a discrete graph representation of large continuous MDPs, where nodes correspond to skill subgoals and the edges to skill policies. The agent constructs this graph during an unsupervised training phase where it interleaves discovering skills and planning using them to gain coverage over ever-increasing portions of the state-space. Given a novel goal at test time, the agent plans with the acquired skill graph to reach a nearby state, then switches to learning to reach the goal. We show that the resulting algorithm, Deep Skill Graphs, outperforms both flat and existing hierarchical reinforcement learning methods on four difficult continuous control tasks.

The posterior over Bayesian neural network (BNN) parameters is extremely high-dimensional and non-convex. For computational reasons, researchers approximate this posterior using inexpensive mini-batch methods such as mean-field variational inference or stochastic-gradient Markov chain Monte Carlo (SGMCMC). To investigate foundational questions in Bayesian deep learning, we instead use full batch Hamiltonian Monte Carlo (HMC) on modern architectures. We show that (1) BNNs can achieve significant performance gains over standard training and deep ensembles; (2) a single long HMC chain can provide a comparable representation of the posterior to multiple shorter chains; (3) in contrast to recent studies, we find posterior tempering is not needed for near-optimal performance, with little evidence for a ``cold posterior'' effect, which we show is largely an artifact of data augmentation; (4) BMA performance is robust to the choice of prior scale, and relatively similar for diagonal Gaussian, mixture of Gaussian, and logistic priors; (5) Bayesian neural networks show surprisingly poor generalization under domain shift; (6) while cheaper alternatives such as deep ensembles and SGMCMC can provide good generalization, their predictive distributions are distinct from HMC. Notably, deep ensemble predictive distributions are similarly close to HMC as standard SGLD, and closer than standard variational inference.

In this paper we cast neural networks defined on graphs as message-passing neural networks (MPNNs) to study the distinguishing power of different classes of such models. We are interested in when certain architectures are able to tell vertices apart based on the feature labels given as input with the graph. We consider two variants of MPNNS: anonymous MPNNs whose message functions depend only on the labels of vertices involved; and degree-aware MPNNs whose message functions can additionally use information regarding the degree of vertices. The former class covers popular graph neural network (GNN) formalisms for which the distinguished power is known. The latter covers graph convolutional networks (GCNs), introduced by Kipf and Welling, for which the distinguishing power was unknown. We obtain lower and upper bounds on the distinguishing power of (anonymous and degree-aware) MPNNs in terms of the distinguishing power of the Weisfeiler-Lehman (WL) algorithm. Our main results imply that (i) the distinguishing power of GCNs is bounded by the WL algorithm, but they may be one step ahead; (ii) the WL algorithm cannot be simulated by ``plain vanilla'' GCNs but the addition of a trade-off parameter between features of the vertex and those of its neighbours (as proposed by Kipf and Welling) resolves this problem.

In this paper, we propose a novel stochastic gradient estimator---ProbAbilistic Gradient Estimator (PAGE)---for nonconvex optimization. PAGE is easy to implement as it is designed via a small adjustment to vanilla SGD: in each iteration, PAGE uses the vanilla minibatch SGD update with probability $p_t$ or reuses the previous gradient with a small adjustment, at a much lower computational cost, with probability $1-p_t$. We give a simple formula for the optimal choice of $p_t$. Moreover, we prove the first tight lower bound $\Omega(n+\frac{\sqrt{n}}{\epsilon^2})$ for nonconvex finite-sum problems, which also leads to a tight lower bound $\Omega(b+\frac{\sqrt{b}}{\epsilon^2})$ for nonconvex online problems, where $b:= \min\{\frac{\sigma^2}{\epsilon^2}, n\}$. Then, we show that PAGE obtains the optimal convergence results $O(n+\frac{\sqrt{n}}{\epsilon^2})$ (finite-sum) and $O(b+\frac{\sqrt{b}}{\epsilon^2})$ (online) matching our lower bounds for both nonconvex finite-sum and online problems. Besides, we also show that for nonconvex functions satisfying the Polyak-\L ojasiewicz (PL) condition, PAGE can automatically switch to a faster linear convergence rate $O(\cdot\log \frac{1}{\epsilon})$. Finally, we conduct several deep learning experiments (e.g., LeNet, VGG, ResNet) on real datasets in PyTorch showing that PAGE not only converges much faster than SGD in training but also achieves the higher test accuracy, validating the optimal theoretical results and confirming the practical superiority of PAGE.

Existing methods for vision-and-language learning typically require designing task-specific architectures and objectives for each task. For example, a multi-label answer classifier for visual question answering, a region scorer for referring expression comprehension, and a language decoder for image captioning, etc. To alleviate these hassles, in this work, we propose a unified framework that learns different tasks in a single architecture with the same language modeling objective, i.e., multimodal conditional text generation, where our models learn to generate labels in text based on the visual and textual inputs. On 7 popular vision-and-language benchmarks, including visual question answering, referring expression comprehension, visual commonsense reasoning, most of which have been previously modeled as discriminative tasks, our generative approach (with a single unified architecture) reaches comparable performance to recent task-specific state-of-the-art vision-and-language models. Moreover, our generative approach shows better generalization ability on questions that have rare answers. Also, we show that our framework allows multi-task learning in a single architecture with a single set of parameters, achieving similar performance to separately optimized single-task models. Our code is publicly available at: https://github.com/j-min/VL-T5

Transformers have been successful for many natural language processing tasks. However, applying transformers to the video domain for tasks such as long-term video generation and scene understanding has remained elusive due to the high computational complexity and the lack of natural tokenization. In this paper, we propose the ObjectCentric Video Transformer (OCVT) which utilizes an object-centric approach for decomposing scenes into tokens suitable for use in a generative video transformer. By factoring the video into objects, our fully unsupervised model is able to learn complex spatio-temporal dynamics of multiple interacting objects in a scene and generate future frames of the video. Our model is also significantly more memory-efficient than pixel-based models and thus able to train on videos of length up to 70 frames with a single 48GB GPU. We compare our model with previous RNN-based approaches as well as other possible video transformer baselines. We demonstrate OCVT performs well when compared to baselines in generating future frames. OCVT also develops useful representations for video reasoning, achieving start-of-the-art performance on the CATER task.

Planning, the ability to analyze the structure of a problem in the large and decompose it into interrelated subproblems, is a hallmark of human intelligence. While deep reinforcement learning (RL) has shown great promise for solving relatively straightforward control tasks, it remains an open problem how to best incorporate planning into existing deep RL paradigms to handle increasingly complex environments. One prominent framework, Model-Based RL, learns a world model and plans using step-by-step virtual rollouts. This type of world model quickly diverges from reality when the planning horizon increases, thus struggling at long-horizon planning. How can we learn world models that endow agents with the ability to do temporally extended reasoning? In this work, we propose to learn graph-structured world models composed of sparse, multi-step transitions. We devise a novel algorithm to learn latent landmarks that are scattered (in terms of reachability) across the goal space as the nodes on the graph. In this same graph, the edges are the reachability estimates distilled from Q-functions. On a variety of high-dimensional continuous control tasks ranging from robotic manipulation to navigation, we demonstrate that our method, named L3P, significantly outperforms prior work, and is oftentimes the only method capable of leveraging both the robustness of model-free RL and generalization of graph-search algorithms. We believe our work is an important step towards scalable planning in reinforcement learning.

Recently, particle-based variational inference (ParVI) methods have gained interest because they can avoid arbitrary parametric assumptions that are common in variational inference. However, many ParVI approaches do not allow arbitrary sampling from the posterior, and the few that do allow such sampling suffer from suboptimality. This work proposes a new method for learning to approximately sample from the posterior distribution. We construct a neural sampler that is trained with the functional gradient of the KL-divergence between the empirical sampling distribution and the target distribution, assuming the gradient resides within a reproducing kernel Hilbert space. Our generative ParVI (GPVI) approach maintains the asymptotic performance of ParVI methods while offering the flexibility of a generative sampler. Through carefully constructed experiments, we show that GPVI outperforms previous generative ParVI methods such as amortized SVGD, and is competitive with ParVI as well as gold-standard approaches like Hamiltonian Monte Carlo for fitting both exactly known and intractable target distributions.

Learning curves model a classifier's test error as a function of the number of training samples. Prior works show that learning curves can be used to select model parameters and extrapolate performance. We investigate how to use learning curves to evaluate design choices, such as pretraining, architecture, and data augmentation. We propose a method to robustly estimate learning curves, abstract their parameters into error and data-reliance, and evaluate the effectiveness of different parameterizations. Our experiments exemplify use of learning curves for analysis and yield several interesting observations.

Reinforcement Learning in large action spaces is a challenging problem. This is especially true for cooperative multi-agent reinforcement learning (MARL), which often requires tractable learning while respecting various constraints like communication budget and information about other agents. In this work, we focus on the fundamental hurdle affecting both value-based and policy-gradient approaches: an exponential blowup of the action space with the number of agents. For value-based methods, it poses challenges in accurately representing the optimal value function for value-based methods, thus inducing suboptimality. For policy gradient methods, it renders the critic ineffective and exacerbates the problem of the lagging critic. We show that from a learning theory perspective, both problems can be addressed by accurately representing the associated action-value function with a low-complexity hypothesis class. This requires accurately modelling the agent interactions in a sample efficient way. To this end, we propose a novel tensorised formulation of the Bellman equation. This gives rise to our method Tesseract, which utilises the view of Q-function seen as a tensor where the modes correspond to action spaces of different agents. Algorithms derived from Tesseract decompose the Q-tensor across the agents and utilise low-rank tensor approximations to model the agent interactions relevant to the task. We provide PAC analysis for Tesseract based algorithms and highlight their relevance to the class of rich observation MDPs. Empirical results in different domains confirm the gains in sample efficiency using Tesseract as supported by the theory.

Adversarial training is among the most effective techniques to improve robustness of models against adversarial perturbations. However, the full effect of this approach on models is not well understood. For example, while adversarial training can reduce the adversarial risk (prediction error against an adversary), it sometimes increase standard risk (generalization error when there is no adversary). In this paper, we focus on \emph{distribution perturbing} adversary framework wherein the adversary can change the test distribution within a neighborhood of the training data distribution. The neighborhood is defined via Wasserstein distance between distributions and the radius of the neighborhood is a measure of adversary's manipulative power. We study the tradeoff between standard risk and adversarial risk and derive the Pareto-optimal tradeoff, achievable over specific classes of models, in the infinite data limit with features dimension kept fixed. We consider three learning settings: 1) Regression with the class of linear models; 2) Binary classification under the Gaussian mixtures data model, with the class of linear classifiers; 3) Regression with the class of random features model (which can be equivalently represented as two-layer neural network with random first-layer weights). We show that a tradeoff between standard and adversarial risk is manifested in all three settings. We further characterize the Pareto-optimal tradeoff curves and discuss how a variety of factors, such as features correlation, adversary's power or the width of two-layer neural network would affect this tradeoff.

Wasserstein Barycenter is a principled approach to represent the weighted mean of a given set of probability distributions, utilizing the geometry induced by optimal transport. In this work, we present a novel scalable algorithm to approximate the Wasserstein Barycenters aiming at high-dimensional applications in machine learning. Our proposed algorithm is based on the Kantorovich dual formulation of the Wasserstein-2 distance as well as a recent neural network architecture, input convex neural network, that is known to parametrize convex functions. The distinguishing features of our method are: i) it only requires samples from the marginal distributions; ii) unlike the existing approaches, it represents the Barycenter with a generative model and can thus generate infinite samples from the barycenter without querying the marginal distributions; iii) it works similar to Generative Adversarial Model in one marginal case. We demonstratethe efficacy of our algorithm by comparing it with the state-of-art methods in multiple experiments.

Efficient gradient computation of the Jacobian determinant term is a core problem in many machine learning settings, and especially so in the normalizing flow framework. Most proposed flow models therefore either restrict to a function class with easy evaluation of the Jacobian determinant, or an efficient estimator thereof. However, these restrictions limit the performance of such density models, frequently requiring significant depth to reach desired performance levels. In this work, we propose \emph{Self Normalizing Flows}, a flexible framework for training normalizing flows by replacing expensive terms in the gradient by learned approximate inverses at each layer. This reduces the computational complexity of each layer's exact update from $\mathcal{O}(D^3)$ to $\mathcal{O}(D^2)$, allowing for the training of flow architectures which were otherwise computationally infeasible, while also providing efficient sampling. We show experimentally that such models are remarkably stable and optimize to similar data likelihood values as their exact gradient counterparts, while training more quickly and surpassing the performance of functionally constrained counterparts.

Projection-free conditional gradient (CG) methods are the algorithms of choice for constrained optimization setups in which projections are often computationally prohibitive but linear optimization over the constraint set remains computationally feasible. Unlike in projection-based methods, globally accelerated convergence rates are in general unattainable for CG. However, a very recent work on Locally accelerated CG (LaCG) has demonstrated that local acceleration for CG is possible for many settings of interest. The main downside of LaCG is that it requires knowledge of the smoothness and strong convexity parameters of the objective function. We remove this limitation by introducing a novel, Parameter-Free Locally accelerated CG (PF-LaCG) algorithm, for which we provide rigorous convergence guarantees. Our theoretical results are complemented by numerical experiments, which demonstrate local acceleration and showcase the practical improvements of PF-LaCG over non-accelerated algorithms, both in terms of iteration count and wall-clock time.

Collecting and aggregating information from several probability measures or histograms is a fundamental task in machine learning. One of the popular solution methods for this task is to compute the barycenter of the probability measures under the Wasserstein metric. However, approximating the Wasserstein barycenter is numerically challenging because of the curse of dimensionality. This paper proposes the projection robust Wasserstein barycenter (PRWB) that has the potential to mitigate the curse of dimensionality, and a relaxed PRWB (RPRWB) model that is computationally more tractable. By combining the iterative Bregman projection algorithm and Riemannian optimization, we propose two algorithms for computing the RPRWB, which is a max-min problem over the Stiefel manifold. The complexity of arithmetic operations of the proposed algorithms for obtaining an $\epsilon$-stationary solution is analyzed. We incorporate the RPRWB into a discrete distribution clustering algorithm, and the numerical results on real text datasets confirm that our RPRWB model helps improve the clustering performance significantly.

Optimal transport (OT) measures distances between distributions in a way that depends on the geometry of the sample space. In light of recent advances in computational OT, OT distances are widely used as loss functions in machine learning. Despite their prevalence and advantages, OT loss functions can be extremely sensitive to outliers. In fact, a single adversarially-picked outlier can increase the standard $W_2$-distance arbitrarily. To address this issue, we propose an outlier-robust formulation of OT. Our formulation is convex but challenging to scale at a first glance. Our main contribution is deriving an \emph{equivalent} formulation based on cost truncation that is easy to incorporate into modern algorithms for computational OT. We demonstrate the benefits of our formulation in mean estimation problems under the Huber contamination model in simulations and outlier detection tasks on real data.

We propose NeRF-VAE, a 3D scene generative model that incorporates geometric structure via Neural Radiance Fields (NeRF) and differentiable volume rendering. In contrast to NeRF, our model takes into account shared structure across scenes, and is able to infer the structure of a novel scene---without the need to re-train---using amortized inference. NeRF-VAE's explicit 3D rendering process further contrasts previous generative models with convolution-based rendering which lacks geometric structure. Our model is a VAE that learns a distribution over radiance fields by conditioning them on a latent scene representation. We show that, once trained, NeRF-VAE is able to infer and render geometrically-consistent scenes from previously unseen 3D environments of synthetic scenes using very few input images. We further demonstrate that NeRF-VAE generalizes well to out-of-distribution cameras, while convolutional models do not. Finally, we introduce and study an attention-based conditioning mechanism of NeRF-VAE's decoder, which improves model performance.

We present a convolution-free approach to video classification built exclusively on self-attention over space and time. Our method, named TimeSformer,'' adapts the standard Transformer architecture to video by enabling spatiotemporal feature learning directly from a sequence of frame-level patches. Our experimental study compares different self-attention schemes and suggests thatdivided attention,'' where temporal attention and spatial attention are separately applied within each block, leads to the best video classification accuracy among the design choices considered. Despite the radically new design, TimeSformer achieves state-of-the-art results on several action recognition benchmarks, including the best reported accuracy on Kinetics-400 and Kinetics-600. Finally, compared to 3D convolutional networks, our model is faster to train, it can achieve dramatically higher test efficiency (at a small drop in accuracy), and it can also be applied to much longer video clips (over one minute long). Code and models are available at: https://github.com/facebookresearch/TimeSformer.

Since the introduction of DQN, a vast majority of reinforcement learning research has focused on reinforcement learning with deep neural networks as function approximators. New methods are typically evaluated on a set of environments that have now become standard, such as Atari 2600 games. While these benchmarks help standardize evaluation, their computational cost has the unfortunate side effect of widening the gap between those with ample access to computational resources, and those without. In this work we argue that, despite the community’s emphasis on large-scale environments, the traditional small-scale environments can still yield valuable scientific insights and can help reduce the barriers to entry for underprivileged communities. To substantiate our claims, we empirically revisit the paper which introduced the Rainbow algorithm [Hessel et al., 2018] and present some new insights into the algorithms used by Rainbow.

Recent developments in the field of model-based RL have proven successful in a range of environments, especially ones where planning is essential. However, such successes have been limited to deterministic fully-observed environments. We present a new approach that handles stochastic and partially-observable environments. Our key insight is to use discrete autoencoders to capture the multiple possible effects of an action in a stochastic environment. We use a stochastic variant of Monte Carlo tree search to plan over both the agent's actions and the discrete latent variables representing the environment's response. Our approach significantly outperforms an offline version of MuZero on a stochastic interpretation of chess where the opponent is considered part of the environment. We also show that our approach scales to DeepMind Lab, a first-person 3D environment with large visual observations and partial observability.

We consider the construction of neural network architectures for data on simplicial complexes. In studying maps on the chain complex of a simplicial complex, we define three desirable properties of a simplicial neural network architecture: namely, permutation equivariance, orientation equivariance, and simplicial awareness. The first two properties respectively account for the fact that the node indexing and the simplex orientations in a simplicial complex are arbitrary. The last property encodes the desirable feature that the output of the neural network depends on the entire simplicial complex and not on a subset of its dimensions. Based on these properties, we propose a simple convolutional architecture, rooted in tools from algebraic topology, for the problem of trajectory prediction, and show that it obeys all three of these properties when an odd, nonlinear activation function is used. We then demonstrate the effectiveness of this architecture in extrapolating trajectories on synthetic and real datasets, with particular emphasis on the gains in generalizability to unseen trajectories.

Decentralization is a promising method of scaling up parallel machine learning systems. In this paper, we provide a tight lower bound on the iteration complexity for such methods in a stochastic non-convex setting. Our lower bound reveals a theoretical gap in known convergence rates of many existing decentralized training algorithms, such as D-PSGD. We prove by construction this lower bound is tight and achievable. Motivated by our insights, we further propose DeTAG, a practical gossip-style decentralized algorithm that achieves the lower bound with only a logarithm gap. Empirically, we compare DeTAG with other decentralized algorithms on image classification tasks, and we show DeTAG enjoys faster convergence compared to baselines, especially on unshuffled data and in sparse networks.

This paper proposes Omnidirectional Representations from Transformers (OMNINET). In OmniNet, instead of maintaining a strictly horizon-tal receptive field, each token is allowed to attend to all tokens in the entire network. This process can also be interpreted as a form of extreme or intensive attention mechanism that has the receptive field of the entire width and depth of the network. To this end, the omnidirectional attention is learned via a meta-learner, which is essentially another self-attention based model. In order to mitigate the computationally expensive costs of full receptive field attention, we leverage efficient self-attention models such as kernel-based, low-rank attention and/or Big Bird as the meta-learner. Extensive experiments are conducted on autoregressive language modeling(LM1B, C4), Machine Translation, Long Range Arena (LRA), and Image Recognition.The experiments show that OmniNet achieves considerable improvements across these tasks, including achieving state-of-the-art performance on LM1B,WMT’14 En-De/En-Fr, and Long Range Arena.Moreover, using omnidirectional representation in Vision Transformers leads to significant improvements on image recognition tasks on both few-shot learning and fine-tuning setups.

We employ constraints to control the parameter space of deep neural networks throughout training. The use of customised, appropriately designed constraints can reduce the vanishing/exploding gradients problem, improve smoothness of classification boundaries, control weight magnitudes and stabilize deep neural networks, and thus enhance the robustness of training algorithms and the generalization capabilities of neural networks. We provide a general approach to efficiently incorporate constraints into a stochastic gradient Langevin framework, allowing enhanced exploration of the loss landscape. We also present specific examples of constrained training methods motivated by orthogonality preservation for weight matrices and explicit weight normalizations. Discretization schemes are provided both for the overdamped formulation of Langevin dynamics and the underdamped form, in which momenta further improve sampling efficiency. These optimisation schemes can be used directly, without needing to adapt neural network architecture design choices or to modify the objective with regularization terms, and see performance improvements in classification tasks.

Realistic use of neural networks often requires adhering to multiple constraints on latency, energy and memory among others. A popular approach to find fitting networks is through constrained Neural Architecture Search (NAS), however, previous methods enforce the constraint only softly. Therefore, the resulting networks do not exactly adhere to the resource constraint and their accuracy is harmed. In this work we resolve this by introducing Hard Constrained diffeRentiable NAS (HardCoRe-NAS), that is based on an accurate formulation of the expected resource requirement and a scalable search method that satisfies the hard constraint throughout the search. Our experiments show that HardCoRe-NAS generates state-of-the-art architectures, surpassing other NAS methods, while strictly satisfying the hard resource constraints without any tuning required.

Attention-based architectures have become ubiquitous in machine learning. Yet, our understanding of the reasons for their effectiveness remains limited. This work proposes a new way to understand self-attention networks: we show that their output can be decomposed into a sum of smaller terms---or paths---each involving the operation of a sequence of attention heads across layers. Using this path decomposition, we prove that self-attention possesses a strong inductive bias towards "token uniformity". Specifically, without skip connections or multi-layer perceptrons (MLPs), the output converges doubly exponentially to a rank-1 matrix. On the other hand, skip connections and MLPs stop the output from degeneration. Our experiments verify the convergence results on standard transformer architectures.

Off-policy evaluation learns a target policy's value with a historical dataset generated by a different behavior policy. In addition to a point estimate, many applications would benefit significantly from having a confidence interval (CI) that quantifies the uncertainty of the point estimate. In this paper, we propose a novel procedure to construct an efficient, robust, and flexible CI on a target policy's value. Our method is justified by theoretical results and numerical experiments. A Python implementation of the proposed procedure is available at https://github.com/ RunzheStat/D2OPE.

Neural differential equations may be trained by backpropagating gradients via the adjoint method, which is another differential equation typically solved using an adaptive-step-size numerical differential equation solver. A proposed step is accepted if its error, \emph{relative to some norm}, is sufficiently small; else it is rejected, the step is shrunk, and the process is repeated. Here, we demonstrate that the particular structure of the adjoint equations makes the usual choices of norm (such as $L^2$) unnecessarily stringent. By replacing it with a more appropriate (semi)norm, fewer steps are unnecessarily rejected and the backpropagation is made faster. This requires only minor code modifications. Experiments on a wide range of tasks---including time series, generative modeling, and physical control---demonstrate a median improvement of 40\% fewer function evaluations. On some problems we see as much as 62\% fewer function evaluations, so that the overall training time is roughly halved.

Temporal-Difference (TD) learning is a general and very useful tool for estimating the value function of a given policy, which in turn is required to find good policies. Generally speaking, TD learning updates states whenever they are visited. When the agent lands in a state, its value can be used to compute the TD-error, which is then propagated to other states. However, it may be interesting, when computing updates, to take into account other information than whether a state is visited or not. For example, some states might be more important than others (such as states which are frequently seen in a successful trajectory). Or, some states might have unreliable value estimates (for example, due to partial observability or lack of data), making their values less desirable as targets. We propose an approach to re-weighting states used in TD updates, both when they are the input and when they provide the target for the update. We prove that our approach converges with linear function approximation and illustrate its desirable empirical behaviour compared to other TD-style methods.

The ability to plan into the future while utilizing only raw high-dimensional observations, such as images, can provide autonomous agents with broad and general capabilities. However, realistic tasks require performing temporally extended reasoning, and cannot be solved with only myopic, short-sighted planning. Recent work in model-based reinforcement learning (RL) has shown impressive results on tasks that require only short-horizon reasoning. In this work, we study how the long-horizon planning abilities can be improved with an algorithm that optimizes over sequences of states, rather than actions, which allows better credit assignment. To achieve this, we draw on the idea of collocation and adapt it to the image-based setting by leveraging probabilistic latent variable models, resulting in an algorithm that optimizes trajectories over latent variables. Our latent collocation method (LatCo) provides a general and effective visual planning approach, and significantly outperforms prior model-based approaches on challenging visual control tasks with sparse rewards and long-term goals. See the videos on the supplementary website \url{https://sites.google.com/view/latco-mbrl/.}

Current graph representation (GR) algorithms require huge demand of human experts in hyperparameter tuning, which significantly limits their practical applications, leading to an urge for automated graph representation without human intervention. Although automated machine learning (AutoML) serves as a good candidate for automatic hyperparameter tuning, little literature has been reported on automated graph presentation learning and the only existing work employs a black-box strategy, lacking insights into explaining the relative importance of different hyperparameters. To address this issue, we study explainable automated graph representation with hyperparameter importance in this paper. We propose an explainable AutoML approach for graph representation (e-AutoGR) which utilizes explainable graph features during performance estimation and learns decorrelated importance weights for different hyperparameters in affecting the model performance through a non-linear decorrelated weighting regression. These learned importance weights can in turn help to provide more insights in hyperparameter search procedure. We theoretically prove the soundness of the decorrelated weighting algorithm. Extensive experiments on real-world datasets demonstrate the superiority of our proposed e-AutoGR model against state-of-the-art methods in terms of both model performance and hyperparameter importance explainability.

Transformer is a ubiquitous model for natural language processing and has attracted wide attentions in computer vision. The attention maps are indispensable for a transformer model to encode the dependencies among input tokens. However, they are learned independently in each layer and sometimes fail to capture precise patterns. In this paper, we propose a novel and generic mechanism based on evolving attention to improve the performance of transformers. On one hand, the attention maps in different layers share common knowledge, thus the ones in preceding layers can instruct the attention in succeeding layers through residual connections. On the other hand, low-level and high-level attentions vary in the level of abstraction, so we adopt convolutional layers to model the evolutionary process of attention maps. The proposed evolving attention mechanism achieves significant performance improvement over various state-of-the-art models for multiple tasks, including image classification, natural language understanding and machine translation.

Vision-and-Language Pre-training (VLP) has improved performance on various joint vision-and-language downstream tasks. Current approaches to VLP heavily rely on image feature extraction processes, most of which involve region supervision (e.g., object detection) and the convolutional architecture (e.g., ResNet). Although disregarded in the literature, we find it problematic in terms of both (1) efficiency/speed, that simply extracting input features requires much more computation than the multimodal interaction steps; and (2) expressive power, as it is upper bounded to the expressive power of the visual embedder and its predefined visual vocabulary. In this paper, we present a minimal VLP model, Vision-and-Language Transformer (ViLT), monolithic in the sense that the processing of visual inputs is drastically simplified to just the same convolution-free manner that we process textual inputs. We show that ViLT is up to tens of times faster than previous VLP models, yet with competitive or better downstream task performance. Our code and pre-trained weights are available at https://github.com/dandelin/vilt.

Modeling distributions on Riemannian manifolds is a crucial component in understanding non-Euclidean data that arises, e.g., in physics and geology. The budding approaches in this space are limited by representational and computational tradeoffs. We propose and study a class of flows that uses convex potentials from Riemannian optimal transport. These are universal and can model distributions on any compact Riemannian manifold without requiring domain knowledge of the manifold to be integrated into the architecture. We demonstrate that these flows can model standard distributions on spheres, and tori, on synthetic and geological data.

In practical applications of iterative first-order optimization, the learning rate schedule remains notoriously difficult to understand and expensive to tune. We demonstrate the presence of these subtleties even in the innocuous case when the objective is a convex quadratic. We reinterpret an iterative algorithm from the numerical analysis literature as what we call the Chebyshev learning rate schedule for accelerating vanilla gradient descent, and show that the problem of mitigating instability leads to a fractal ordering of step sizes. We provide some experiments to challenge conventional beliefs about stable learning rates in deep learning: the fractal schedule enables training to converge with locally unstable updates which make negative progress on the objective.

Existing evaluation suites for multi-agent reinforcement learning (MARL) do not assess generalization to novel situations as their primary objective (unlike supervised learning benchmarks). Our contribution, Melting Pot, is a MARL evaluation suite that fills this gap and uses reinforcement learning to reduce the human labor required to create novel test scenarios. This works because one agent's behavior constitutes (part of) another agent's environment. To demonstrate scalability, we have created over 80 unique test scenarios covering a broad range of research topics such as social dilemmas, reciprocity, resource sharing, and task partitioning. We apply these test scenarios to standard MARL training algorithms, and demonstrate how Melting Pot reveals weaknesses not apparent from training performance alone.

Recent results in supervised learning suggest that while overparameterized models have the capacity to overfit, they in fact generalize quite well. We ask whether the same phenomenon occurs for offline contextual bandits. Our results are mixed. Value-based algorithms benefit from the same generalization behavior as overparameterized supervised learning, but policy-based algorithms do not. We show that this discrepancy is due to the \emph{action-stability} of their objectives. An objective is action-stable if there exists a prediction (action-value vector or action distribution) which is optimal no matter which action is observed. While value-based objectives are action-stable, policy-based objectives are unstable. We formally prove upper bounds on the regret of overparameterized value-based learning and lower bounds on the regret for policy-based algorithms. In our experiments with large neural networks, this gap between action-stable value-based objectives and unstable policy-based objectives leads to significant performance differences.

This paper bridges discrete and continuous optimization approaches for decomposable submodular function minimization, in both the standard and parametric settings. We provide improved running times for this problem by reducing it to a number of calls to a maximum flow oracle. When each function in the decomposition acts on O(1) elements of the ground set V and is polynomially bounded, our running time is up to polylogarithmic factors equal to that of solving maximum flow in a sparse graph with O(|V|) vertices and polynomial integral capacities. We achieve this by providing a simple iterative method which can optimize to high precision any convex function defined on the submodular base polytope, provided we can efficiently minimize it on the base polytope corresponding to the cut function of a certain graph that we construct. We solve this minimization problem by lifting the solutions of a parametric cut problem, which we obtain via a new efficient combinatorial reduction to maximum flow. This reduction is of independent interest and implies some previously unknown bounds for the parametric minimum s,t-cut problem in multiple settings.

The lottery ticket hypothesis (LTH) reveals the existence of winning tickets (sparse but critical subnetworks) for dense networks, that can be trained in isolation from random initialization to match the latter's accuracies. However, finding winning tickets requires burdensome computations in the train-prune-retrain process, especially on large-scale datasets (e.g., ImageNet), restricting their practical benefits. This paper explores a new perspective on finding lottery tickets more efficiently, by doing so only with a specially selected subset of data, called Pruning-Aware Critical set (PrAC set), rather than using the full training set. The concept of PrAC set was inspired by the recent observation, that deep networks have samples that are either hard to memorize during training, or easy to forget during pruning. A PrAC set is thus hypothesized to capture those most challenging and informative examples for the dense model. We observe that a high-quality winning ticket can be found with training and pruning the dense network on the very compact PrAC set, which can substantially save training iterations for the ticket finding process. Extensive experiments validate our proposal across diverse datasets and network architectures. Specifically, on CIFAR-10, CIFAR-100, and Tiny ImageNet, we locate effective PrAC sets at 35.32%~78.19% of their training set sizes. On top of them, we can obtain the same competitive winning tickets for the corresponding dense networks, yet saving up to 82.85%~92.77%, 63.54%~74.92%, and 76.14%~86.56% training iterations, respectively. Crucially, we show that a PrAC set found is reusable across different network architectures, which can amortize the extra cost of finding PrAC sets, yielding a practical regime for efficient lottery ticket finding.

With graphs rapidly growing in size and deeper graph neural networks (GNNs) emerging, the training and inference of GNNs become increasingly expensive. Existing network weight pruning algorithms cannot address the main space and computational bottleneck in GNNs, caused by the size and connectivity of the graph. To this end, this paper first presents a unified GNN sparsification (UGS) framework that simultaneously prunes the graph adjacency matrix and the model weights, for effectively accelerating GNN inference on large-scale graphs. Leveraging this new tool, we further generalize the recently popular lottery ticket hypothesis to GNNs for the first time, by defining a graph lottery ticket (GLT) as a pair of core sub-dataset and sparse sub-network, which can be jointly identified from the original GNN and the full dense graph by iteratively applying UGS. Like its counterpart in convolutional neural networks, GLT can be trained in isolation to match the performance of training with the full model and graph, and can be drawn from both randomly initialized and self-supervised pre-trained GNNs. Our proposal has been experimentally verified across various GNN architectures and diverse tasks, on both small-scale graph datasets (Cora, Citeseer and PubMed), and large-scale datasets from the challenging Open Graph Benchmark (OGB). Specifically, for node classification, our found GLTs achieve the same accuracies with 20%~98% MACs saving on small graphs and 25%~85% MACs saving on large ones. For link prediction, GLTs lead to 48%~97% and 70% MACs saving on small and large graph datasets, respectively, without compromising predictive performance. Codes are at https://github.com/VITA-Group/Unified-LTH-GNN.

The success of Convolutional Neural Networks (CNNs) in computer vision is mainly driven by their strong inductive bias, which is strong enough to allow CNNs to solve vision-related tasks with random weights, meaning without learning. Similarly, Long Short-Term Memory (LSTM) has a strong inductive bias towards storing information over time. However, many real-world systems are governed by conservation laws, which lead to the redistribution of particular quantities — e.g.in physical and economical systems. Our novel Mass-Conserving LSTM (MC-LSTM) adheres to these conservation laws by extending the inductive bias of LSTM to model the redistribution of those stored quantities. MC-LSTMs set a new state-of-the-art for neural arithmetic units at learning arithmetic operations, such as addition tasks,which have a strong conservation law, as the sum is constant over time. Further, MC-LSTM is applied to traffic forecasting, modeling a pendulum, and a large benchmark dataset in hydrology, where it sets a new state-of-the-art for predicting peak flows. In the hydrology example, we show that MC-LSTM states correlate with real world processes and are therefore interpretable.

Denoising diffusion probabilistic models (DDPM) are a class of generative models which have recently been shown to produce excellent samples. We show that with a few simple modifications, DDPMs can also achieve competitive log-likelihoods while maintaining high sample quality. Additionally, we find that learning variances of the reverse diffusion process allows sampling with an order of magnitude fewer forward passes with a negligible difference in sample quality, which is important for the practical deployment of these models. We additionally use precision and recall to compare how well DDPMs and GANs cover the target distribution. Finally, we show that the sample quality and likelihood of these models scale smoothly with model capacity and training compute, making them easily scalable. We release our code and pre-trained models at https://github.com/openai/improved-diffusion.

Batch policy optimization considers leveraging existing data for policy construction before interacting with an environment. Although interest in this problem has grown significantly in recent years, its theoretical foundations remain under-developed. To advance the understanding of this problem, we provide three results that characterize the limits and possibilities of batch policy optimization in the finite-armed stochastic bandit setting. First, we introduce a class of confidence-adjusted index algorithms that unifies optimistic and pessimistic principles in a common framework, which enables a general analysis. For this family, we show that any confidence-adjusted index algorithm is minimax optimal, whether it be optimistic, pessimistic or neutral. Our analysis reveals that instance-dependent optimality, commonly used to establish optimality of on-line stochastic bandit algorithms, cannot be achieved by any algorithm in the batch setting. In particular, for any algorithm that performs optimally in some environment, there exists another environment where the same algorithm suffers arbitrarily larger regret. Therefore, to establish a framework for distinguishing algorithms, we introduce a new weighted-minimax criterion that considers the inherent difficulty of optimal value prediction. We demonstrate how this criterion can be used to justify commonly used pessimistic principles for batch policy optimization.

Optimal transport (OT) is a widely used technique for distribution alignment, with applications throughout the machine learning, graphics, and vision communities. Without any additional structural assumptions on transport, however, OT can be fragile to outliers or noise, especially in high dimensions. Here, we introduce Latent Optimal Transport (LOT), a new approach for OT that simultaneously learns low-dimensional structure in data while leveraging this structure to solve the alignment task. The idea behind our approach is to learn two sets of ``anchors'' that constrain the flow of transport between a source and target distribution. In both theoretical and empirical studies, we show that LOT regularizes the rank of transport and makes it more robust to outliers and the sampling density. We show that by allowing the source and target to have different anchors, and using LOT to align the latent spaces between anchors, the resulting transport plan has better structural interpretability and highlights connections between both the individual data points and the local geometry of the datasets.

We present a message passing method for 0–1 integer linear programs. Our algorithm is based on a decomposition of the original problem into subproblems that are represented as binary deci- sion diagrams. The resulting Lagrangean dual is solved iteratively by a series of efficient block coordinate ascent steps. Our method has linear iteration complexity in the size of the decomposi- tion and can be effectively parallelized. The char- acteristics of our approach are desirable towards solving ever larger problems arising in structured prediction. We present experimental results on combinatorial problems from MAP inference for Markov Random Fields, quadratic assignment, discrete tomography and cell tracking for develop- mental biology and show promising performance.

We propose a distributed cubic regularization of the Newton method for solving (constrained) empirical risk minimization problems over a network of agents, modeled as undirected graph. The algorithm employs an inexact, preconditioned Newton step at each agent's side: the gradient of the centralized loss is iteratively estimated via a gradient-tracking consensus mechanism and the Hessian is subsampled over the local data sets. No Hessian matrices are exchanged over the network. We derive global complexity bounds for convex and strongly convex losses. Our analysis reveals an interesting interplay between sample and iteration/communication complexity: statistically accurate solutions are achievable in roughly the same number of iterations of the centralized cubic Newton, with a communication cost per iteration of the order of $\widetilde{\mathcal{O}}\big(1/\sqrt{1-\rho}\big)$, where $\rho$ characterizes the connectivity of the network. This represents a significant improvement with respect to existing, statistically oblivious, distributed Newton-based methods over networks.

Graph neural networks (GNNs) can process graphs of different sizes, but their ability to generalize across sizes, specifically from small to large graphs, is still not well understood. In this paper, we identify an important type of data where generalization from small to large graphs is challenging: graph distributions for which the local structure depends on the graph size. This effect occurs in multiple important graph learning domains, including social and biological networks. We first prove that when there is a difference between the local structures, GNNs are not guaranteed to generalize across sizes: there are "bad" global minima that do well on small graphs but fail on large graphs. We then study the size-generalization problem empirically and demonstrate that when there is a discrepancy in local structure, GNNs tend to converge to non-generalizing solutions. Finally, we suggest two approaches for improving size generalization, motivated by our findings. Notably, we propose a novel Self-Supervised Learning (SSL) task aimed at learning meaningful representations of local structures that appear in large graphs. Our SSL task improves classification accuracy on several popular datasets.

Several recent applications of optimal transport (OT) theory to machine learning have relied on regularization, notably entropy and the Sinkhorn algorithm. Because matrix-vector products are pervasive in the Sinkhorn algorithm, several works have proposed to \textit{approximate} kernel matrices appearing in its iterations using low-rank factors. Another route lies instead in imposing low-nonnegative rank constraints on the feasible set of couplings considered in OT problems, with no approximations on cost nor kernel matrices. This route was first explored by~\citet{forrow2018statistical}, who proposed an algorithm tailored for the squared Euclidean ground cost, using a proxy objective that can be solved through the machinery of regularized 2-Wasserstein barycenters. Building on this, we introduce in this work a generic approach that aims at solving, in full generality, the OT problem under low-nonnegative rank constraints with arbitrary costs. Our algorithm relies on an explicit factorization of low-rank couplings as a product of \textit{sub-coupling} factors linked by a common marginal; similar to an NMF approach, we alternatively updates these factors. We prove the non-asymptotic stationary convergence of this algorithm and illustrate its efficiency on benchmark experiments.

Model-based Reinforcement Learning (RL) is a popular learning paradigm due to its potential sample efficiency compared to model-free RL. However, existing empirical model-based RL approaches lack the ability to explore. This work studies a computationally and statistically efficient model-based algorithm for both Kernelized Nonlinear Regulators (KNR) and linear Markov Decision Processes (MDPs). For both models, our algorithm guarantees polynomial sample complexity and only uses access to a planning oracle. Experimentally, we first demonstrate the flexibility and the efficacy of our algorithm on a set of exploration challenging control tasks where existing empirical model-based RL approaches completely fail. We then show that our approach retains excellent performance even in common dense reward control benchmarks that do not require heavy exploration.

Unsupervised protein language models trained across millions of diverse sequences learn structure and function of proteins. Protein language models studied to date have been trained to perform inference from individual sequences. The longstanding approach in computational biology has been to make inferences from a family of evolutionarily related sequences by fitting a model to each family independently. In this work we combine the two paradigms. We introduce a protein language model which takes as input a set of sequences in the form of a multiple sequence alignment. The model interleaves row and column attention across the input sequences and is trained with a variant of the masked language modeling objective across many protein families. The performance of the model surpasses current state-of-the-art unsupervised structure learning methods by a wide margin, with far greater parameter efficiency than prior state-of-the-art protein language models.

There has been a surge of interest in continual learning and federated learning, both of which are important in deep neural networks in real-world scenarios. Yet little research has been done regarding the scenario where each client learns on a sequence of tasks from a private local data stream. This problem of federated continual learning poses new challenges to continual learning, such as utilizing knowledge from other clients, while preventing interference from irrelevant knowledge. To resolve these issues, we propose a novel federated continual learning framework, Federated Weighted Inter-client Transfer (FedWeIT), which decomposes the network weights into global federated parameters and sparse task-specific parameters, and each client receives selective knowledge from other clients by taking a weighted combination of their task-specific parameters. FedWeIT minimizes interference between incompatible tasks, and also allows positive knowledge transfer across clients during learning. We validate our FedWeIT against existing federated learning and continual learning methods under varying degrees of task similarity across clients, and our model significantly outperforms them with a large reduction in the communication cost.

The variance-stabilizing transformation (VST) problem is to transform heteroscedastic data to homoscedastic data so that they are more tractable for subsequent analysis. However, most of the existing approaches focus on finding an analytical solution for a certain parametric distribution, which severely limits the applications, because simple distributions cannot faithfully describe the real data while more complicated distributions cannot be analytically solved. In this paper, we converted the VST problem into a convex optimization problem, which can always be efficiently solved, identified the specific structure of the convex problem, which further improved the efficiency of the proposed algorithm, and showed that any finite discrete distributions and the discretized version of any continuous distributions from real data can be variance-stabilized in an easy and nonparametric way. We demonstrated the new approach on bioimaging data and achieved superior performance compared to peer algorithms in terms of not only the variance homoscedasticity but also the impact on subsequent analysis such as denoising. Source codes are available at https://github.com/yu-lab-vt/ConvexVST.

This paper considers decentralized stochastic optimization over a network of $n$ nodes, where each node possesses a smooth non-convex local cost function and the goal of the networked nodes is to find an $\epsilon$-accurate first-order stationary point of the sum of the local costs. We focus on an online setting, where each node accesses its local cost only by means of a stochastic first-order oracle that returns a noisy version of the exact gradient. In this context, we propose a novel single-loop decentralized hybrid variance-reduced stochastic gradient method, called GT-HSGD, that outperforms the existing approaches in terms of both the oracle complexity and practical implementation. The GT-HSGD algorithm implements specialized local hybrid stochastic gradient estimators that are fused over the network to track the global gradient. Remarkably, GT-HSGD achieves a network topology-independent oracle complexity of $O(n^{-1}\epsilon^{-3})$ when the required error tolerance $\epsilon$ is small enough, leading to a linear speedup with respect to the centralized optimal online variance-reduced approaches that operate on a single node. Numerical experiments are provided to illustrate our main technical results.

A black-box spectral method is introduced for evaluating the adversarial robustness of a given machine learning (ML) model. Our approach, named SPADE, exploits bijective distance mapping between the input/output graphs constructed for approximating the manifolds corresponding to the input/output data. By leveraging the generalized Courant-Fischer theorem, we propose a SPADE score for evaluating the adversarial robustness of a given model, which is proved to be an upper bound of the best Lipschitz constant under the manifold setting. To reveal the most non-robust data samples highly vulnerable to adversarial attacks, we develop a spectral graph embedding procedure leveraging dominant generalized eigenvectors. This embedding step allows assigning each data point a robustness score that can be further harnessed for more effective adversarial training of ML models. Our experiments show promising empirical results for neural networks trained with the MNIST and CIFAR-10 data sets.

This paper presents a new approach for assembling graph neural networks based on framelet transforms. The latter provides a multi-scale representation for graph-structured data. We decompose an input graph into low-pass and high-pass frequencies coefficients for network training, which then defines a framelet-based graph convolution. The framelet decomposition naturally induces a graph pooling strategy by aggregating the graph feature into low-pass and high-pass spectra, which considers both the feature values and geometry of the graph data and conserves the total information. The graph neural networks with the proposed framelet convolution and pooling achieve state-of-the-art performance in many node and graph prediction tasks. Moreover, we propose shrinkage as a new activation for the framelet convolution, which thresholds high-frequency information at different scales. Compared to ReLU, shrinkage activation improves model performance on denoising and signal compression: noises in both node and structure can be significantly reduced by accurately cutting off the high-pass coefficients from framelet decomposition, and the signal can be compressed to less than half its original size with well-preserved prediction performance.

We develop several new communication-efficient second-order methods for distributed optimization. Our first method, NEWTON-STAR, is a variant of Newton's method from which it inherits its fast local quadratic rate. However, unlike Newton's method, NEWTON-STAR enjoys the same per iteration communication cost as gradient descent. While this method is impractical as it relies on the use of certain unknown parameters characterizing the Hessian of the objective function at the optimum, it serves as the starting point which enables us to design practical variants thereof with strong theoretical guarantees. In particular, we design a stochastic sparsification strategy for learning the unknown parameters in an iterative fashion in a communication efficient manner. Applying this strategy to NEWTON-STAR leads to our next method, NEWTON-LEARN, for which we prove local linear and superlinear rates independent of the condition number. When applicable, this method can have dramatically superior convergence behavior when compared to state-of-the-art methods. Finally, we develop a globalization strategy using cubic regularization which leads to our next method, CUBIC-NEWTON-LEARN, for which we prove global sublinear and linear convergence rates, and a fast superlinear rate. Our results are supported with experimental results on real datasets, and show several orders of magnitude improvement on baseline and state-of-the-art methods in terms of communication complexity.

Bayesian optimization (BO) is among the most effective and widely-used blackbox optimization methods. BO proposes solutions according to an explore-exploit trade-off criterion encoded in an acquisition function, many of which are computed from the posterior predictive of a probabilistic surrogate model. Prevalent among these is the expected improvement (EI). The need to ensure analytical tractability of the predictive often poses limitations that can hinder the efficiency and applicability of BO. In this paper, we cast the computation of EI as a binary classification problem, building on the link between class-probability estimation and density-ratio estimation, and the lesser-known link between density-ratios and EI. By circumventing the tractability constraints, this reformulation provides numerous advantages, not least in terms of expressiveness, versatility, and scalability.

Ratio maximization has applications in areas as diverse as finance, reward shaping for reinforcement learning (RL), and the development of safe artificial intelligence, yet there has been very little exploration of RL algorithms for ratio maximization. This paper addresses this deficiency by introducing two new, model-free RL algorithms for solving cost-aware Markov decision processes, where the goal is to maximize the ratio of long-run average reward to long-run average cost. The first algorithm is a two-timescale scheme based on relative value iteration (RVI) Q-learning and the second is an actor-critic scheme. The paper proves almost sure convergence of the former to the globally optimal solution in the tabular case and almost sure convergence of the latter under linear function approximation for the critic. Unlike previous methods, the two algorithms provably converge for general reward and cost functions under suitable conditions. The paper also provides empirical results demonstrating promising performance and lending strong support to the theoretical results.

Group equivariant neural networks are used as building blocks of group invariant neural networks, which have been shown to improve generalisation performance and data efficiency through principled parameter sharing. Such works have mostly focused on group equivariant convolutions, building on the result that group equivariant linear maps are necessarily convolutions. In this work, we extend the scope of the literature to self-attention, that is emerging as a prominent building block of deep learning models. We propose the LieTransformer, an architecture composed of LieSelfAttention layers that are equivariant to arbitrary Lie groups and their discrete subgroups. We demonstrate the generality of our approach by showing experimental results that are competitive to baseline methods on a wide range of tasks: shape counting on point clouds, molecular property regression and modelling particle trajectories under Hamiltonian dynamics.

We consider Bayesian optimization in settings where observations can be adversarially biased, for example by an uncontrolled hidden confounder. Our first contribution is a reduction of the confounded setting to the dueling bandit model. Then we propose a novel approach for dueling bandits based on information-directed sampling (IDS). Thereby, we obtain the first efficient kernelized algorithm for dueling bandits that comes with cumulative regret guarantees. Our analysis further generalizes a previously proposed semi-parametric linear bandit model to non-linear reward functions, and uncovers interesting links to doubly-robust estimation.

The pairwise interaction paradigm of graph machine learning has predominantly governed the modelling of relational systems. However, graphs alone cannot capture the multi-level interactions present in many complex systems and the expressive power of such schemes was proven to be limited. To overcome these limitations, we propose Message Passing Simplicial Networks (MPSNs), a class of models that perform message passing on simplicial complexes (SCs). To theoretically analyse the expressivity of our model we introduce a Simplicial Weisfeiler-Lehman (SWL) colouring procedure for distinguishing non-isomorphic SCs. We relate the power of SWL to the problem of distinguishing non-isomorphic graphs and show that SWL and MPSNs are strictly more powerful than the WL test and not less powerful than the 3-WL test. We deepen the analysis by comparing our model with traditional graph neural networks (GNNs) with ReLU activations in terms of the number of linear regions of the functions they can represent. We empirically support our theoretical claims by showing that MPSNs can distinguish challenging strongly regular graphs for which GNNs fail and, when equipped with orientation equivariant layers, they can improve classification accuracy in oriented SCs compared to a GNN baseline.

As humans, our goals and our environment are persistently changing throughout our lifetime based on our experiences, actions, and internal and external drives. In contrast, typical reinforcement learning problem set-ups consider decision processes that are stationary across episodes. Can we develop reinforcement learning algorithms that can cope with the persistent change in the former, more realistic problem settings? While on-policy algorithms such as policy gradients in principle can be extended to non-stationary settings, the same cannot be said for more efficient off-policy algorithms that replay past experiences when learning. In this work, we formalize this problem setting, and draw upon ideas from the online learning and probabilistic inference literature to derive an off-policy RL algorithm that can reason about and tackle such lifelong non-stationarity. Our method leverages latent variable models to learn a representation of the environment from current and past experiences, and performs off-policy RL with this representation. We further introduce several simulation environments that exhibit lifelong non-stationarity, and empirically find that our approach substantially outperforms approaches that do not reason about environment shift.

We propose a novel method for exploring how neurons within neural networks interact. In particular, we consider activation values of a network for given data, and propose to mine noise-robust rules of the form X → Y , where X and Y are sets of neurons in different layers. We identify the best set of rules by the Minimum Description Length Principle as the rules that together are most descriptive of the activation data. To learn good rule sets in practice, we propose the unsupervised ExplaiNN algorithm. Extensive evaluation shows that the patterns it discovers give clear insight in how networks perceive the world: they identify shared, respectively class-specific traits, compositionality within the network, as well as locality in convolutional layers. Moreover, these patterns are not only easily interpretable, but also supercharge prototyping as they identify which groups of neurons to consider in unison.

Variational data assimilation optimizes for an initial state of a dynamical system such that its evolution fits observational data. The physical model can subsequently be evolved into the future to make predictions. This principle is a cornerstone of large scale forecasting applications such as numerical weather prediction. As such, it is implemented in current operational systems of weather forecasting agencies across the globe. However, finding a good initial state poses a difficult optimization problem in part due to the non-invertible relationship between physical states and their corresponding observations. We learn a mapping from observational data to physical states and show how it can be used to improve optimizability. We employ this mapping in two ways: to better initialize the non-convex optimization problem, and to reformulate the objective function in better behaved physics space instead of observation space. Our experimental results for the Lorenz96 model and a two-dimensional turbulent fluid flow demonstrate that this procedure significantly improves forecast quality for chaotic systems.

A wide range of optimization problems arising in machine learning can be solved by gradient descent algorithms, and a central question in this area is how to efficiently compress a large-scale dataset so as to reduce the computational complexity. Coreset is a popular data compression technique that has been extensively studied before. However, most of existing coreset methods are problem-dependent and cannot be used as a general tool for a broader range of applications. A key obstacle is that they often rely on the pseudo-dimension and total sensitivity bound that can be very high or hard to obtain. In this paper, based on the locality'' property of gradient descent algorithms, we propose a new framework, termedsequential coreset'', which effectively avoids these obstacles. Moreover, our method is particularly suitable for sparse optimization whence the coreset size can be further reduced to be only poly-logarithmically dependent on the dimension. In practice, the experimental results suggest that our method can save a large amount of running time compared with the baseline algorithms.

Various machine learning tasks, from generative modeling to domain adaptation, revolve around the concept of dataset transformation and manipulation. While various methods exist for transforming unlabeled datasets, principled methods to do so for labeled (e.g., classification) datasets are missing. In this work, we propose a novel framework for dataset transformation, which we cast as optimization over data-generating joint probability distributions. We approach this class of problems through Wasserstein gradient flows in probability space, and derive practical and efficient particle-based methods for a flexible but well-behaved class of objective functions. Through various experiments, we show that this framework can be used to impose constraints on classification datasets, adapt them for transfer learning, or to re-purpose fixed or black-box models to classify —with high accuracy— previously unseen datasets.

Datacenter vision systems widely use small, specialized convolutional neural networks (CNNs) trained on specific tasks for high-throughput inference. These settings employ accelerators with massive computational capacity, but which specialized CNNs underutilize due to having low arithmetic intensity. This results in suboptimal application-level throughput and poor returns on accelerator investment. Increasing batch size is the only known way to increase both application-level throughput and accelerator utilization for inference, but yields diminishing returns; specialized CNNs poorly utilize accelerators even with large batch size. We propose FoldedCNNs, a new approach to CNN design that increases inference throughput and utilization beyond large batch size. FoldedCNNs rethink the structure of inputs and layers of specialized CNNs to boost arithmetic intensity: in FoldedCNNs, f images with C channels each are concatenated into a single input with fC channels and jointly classified by a wider CNN. Increased arithmetic intensity in FoldedCNNs increases the throughput and GPU utilization of specialized CNN inference by up to 2.5x and 2.8x, with accuracy close to the original CNN in most cases.

Current reinforcement learning (RL) methods use simulation models as simple black-box oracles. In this paper, with the goal of improving the performance exhibited by RL algorithms, we explore a systematic way of leveraging the additional information provided by an emerging class of differentiable simulators. Building on concepts established by Deterministic Policy Gradients (DPG) methods, the neural network policies learned with our approach represent deterministic actions. In a departure from standard methodologies, however, learning these policies does not hinge on approximations of the value function that must be learned concurrently in an actor-critic fashion. Instead, we exploit differentiable simulators to directly compute the analytic gradient of a policy's value function with respect to the actions it outputs. This, in turn, allows us to efficiently perform locally optimal policy improvement iterations. Compared against other state-of-the-art RL methods, we show that with minimal hyper-parameter tuning our approach consistently leads to better asymptotic behavior across a set of payload manipulation tasks that demand a high degree of accuracy and precision.

Structured nonsmooth convex finite-sum optimization appears in many machine learning applications, including support vector machines and least absolute deviation. For the primal-dual formulation of this problem, we propose a novel algorithm called \emph{Variance Reduction via Primal-Dual Accelerated Dual Averaging (\vrpda)}. In the nonsmooth and general convex setting, \vrpda~has the overall complexity $O(nd\log\min \{1/\epsilon, n\} + d/\epsilon )$ in terms of the primal-dual gap, where $n$ denotes the number of samples, $d$ the dimension of the primal variables, and $\epsilon$ the desired accuracy. In the nonsmooth and strongly convex setting, the overall complexity of \vrpda~becomes $O(nd\log\min\{1/\epsilon, n\} + d/\sqrt{\epsilon})$ in terms of both the primal-dual gap and the distance between iterate and optimal solution. Both these results for \vrpda~improve significantly on state-of-the-art complexity estimates---which are $O(nd\log \min\{1/\epsilon, n\} + \sqrt{n}d/\epsilon)$ for the nonsmooth and general convex setting and $O(nd\log \min\{1/\epsilon, n\} + \sqrt{n}d/\sqrt{\epsilon})$ for the nonsmooth and strongly convex setting---with a simpler and more straightforward algorithm and analysis. Moreover, both complexities are better than \emph{lower} bounds for general convex finite-sum optimization, because our approach makes use of additional, commonly occurring structure. Numerical experiments reveal competitive performance of \vrpda~compared to state-of-the-art approaches.

Symbolic equations are at the core of scientific discovery. The task of discovering the underlying equation from a set of input-output pairs is called symbolic regression. Traditionally, symbolic regression methods use hand-designed strategies that do not improve with experience. In this paper, we introduce the first symbolic regression method that leverages large scale pre-training. We procedurally generate an unbounded set of equations, and simultaneously pre-train a Transformer to predict the symbolic equation from a corresponding set of input-output-pairs. At test time, we query the model on a new set of points and use its output to guide the search for the equation. We show empirically that this approach can re-discover a set of well-known physical equations, and that it improves over time with more data and compute.

Variational autoencoders with deep hierarchies of stochastic layers have been known to suffer from the problem of posterior collapse, where the top layers fall back to the prior and become independent of input. We suggest that the hierarchical VAE objective explicitly includes the variance of the function parameterizing the mean and variance of the latent Gaussian distribution which itself is often a high variance function. Building on this we generalize VAE neural networks by incorporating a smoothing parameter motivated by Gaussian analysis to reduce higher frequency components and consequently the variance in parameterizing functions and show that this can help to solve the problem of posterior collapse. We further show that under such smoothing the VAE loss exhibits a phase transition, where the top layer KL divergence sharply drops to zero at a critical value of the smoothing parameter that is similar for the same model across datasets. We validate the phenomenon across model configurations and datasets.

We consider the optimal approximate posterior over the top-layer weights in a Bayesian neural network for regression, and show that it exhibits strong dependencies on the lower-layer weights. We adapt this result to develop a correlated approximate posterior over the weights at all layers in a Bayesian neural network. We extend this approach to deep Gaussian processes, unifying inference in the two model classes. Our approximate posterior uses learned "global'' inducing points, which are defined only at the input layer and propagated through the network to obtain inducing inputs at subsequent layers. By contrast, standard, "local'', inducing point methods from the deep Gaussian process literature optimise a separate set of inducing inputs at every layer, and thus do not model correlations across layers. Our method gives state-of-the-art performance for a variational Bayesian method, without data augmentation or tempering, on CIFAR-10 of 86.7%, which is comparable to SGMCMC without tempering but with data augmentation (88% in Wenzel et al. 2020).

One of the fascinating properties of deep learning is the ability of the network to reveal the underlying factors characterizing elements in datasets of different types. Autoencoders represent an effective approach for computing these factors. Autoencoders have been studied in the context of enabling interpolation between data points by decoding convex combinations of latent vectors. However, this interpolation often leads to artifacts or produces unrealistic results during reconstruction. We argue that these incongruities are due to the structure of the latent space and to the fact that such naively interpolated latent vectors deviate from the data manifold. In this paper, we propose a regularization technique that shapes the latent representation to follow a manifold that is consistent with the training images and that forces the manifold to be smooth and locally convex. This regularization not only enables faithful interpolation between data points, as we show herein but can also be used as a general regularization technique to avoid overfitting or to produce new samples for data augmentation.

Strategic classification regards the problem of learning in settings where users can strategically modify their features to improve outcomes. This setting applies broadly, and has received much recent attention. But despite its practical significance, work in this space has so far been predominantly theoretical. In this paper we present a learning framework for strategic classification that is practical. Our approach directly minimizes the ``strategic'' empirical risk, which we achieve by differentiating through the strategic response of users. This provides flexibility that allows us to extend beyond the original problem formulation and towards more realistic learning scenarios. A series of experiments demonstrates the effectiveness of our approach on various learning settings.

Neural network pruning techniques reduce the number of parameters without compromising predicting ability of a network. Many algorithms have been developed for pruning both over-parameterized fully-connected networks (FCN) and convolutional neural networks (CNN), but analytical studies of capabilities and compression ratios of such pruned sub-networks are lacking. We theoretically study the performance of two pruning techniques (random and magnitude-based) on FCN and CNN. Given a target network, we provide a universal approach to bound the gap between a pruned and the target network in a probabilistic sense, which is the first study of this nature. The results establish that there exist pruned networks with expressive power within any specified bound from the target network and with a significant compression ratio.

Linear interpolation between initial neural network parameters and converged parameters after training with stochastic gradient descent (SGD) typically leads to a monotonic decrease in the training objective. This Monotonic Linear Interpolation (MLI) property, first observed by Goodfellow et al. 2014, persists in spite of the non-convex objectives and highly non-linear training dynamics of neural networks. Extending this work, we evaluate several hypotheses for this property that, to our knowledge, have not yet been explored. Using tools from differential geometry, we draw connections between the interpolated paths in function space and the monotonicity of the network --- providing sufficient conditions for the MLI property under mean squared error. While the MLI property holds under various settings (e.g., network architectures and learning problems), we show in practice that networks violating the MLI property can be produced systematically, by encouraging the weights to move far from initialization. The MLI property raises important questions about the loss landscape geometry of neural networks and highlights the need to further study their global properties.

We consider the problem of controlling a partially-observed dynamic process on a graph by a limited number of interventions. This problem naturally arises in contexts such as scheduling virus tests to curb an epidemic; targeted marketing in order to promote a product; and manually inspecting posts to detect fake news spreading on social networks.

We formulate this setup as a sequential decision problem over a temporal graph process. In face of an exponential state space, combinatorial action space and partial observability, we design a novel tractable scheme to control dynamical processes on temporal graphs. We successfully apply our approach to two popular problems that fall into our framework: prioritizing which nodes should be tested in order to curb the spread of an epidemic, and influence maximization on a graph.

We study a stochastic program where the probability distribution of the uncertain problem parameters is unknown and only indirectly observed via finitely many correlated samples generated by an unknown Markov chain with $d$ states. We propose a data-driven distributionally robust optimization model to estimate the problem's objective function and optimal solution. By leveraging results from large deviations theory, we derive statistical guarantees on the quality of these estimators. The underlying worst-case expectation problem is nonconvex and involves $\mathcal O(d^2)$ decision variables. Thus, it cannot be solved efficiently for large $d$. By exploiting the structure of this problem, we devise a customized Frank-Wolfe algorithm with convex direction-finding subproblems of size $\mathcal O(d)$. We prove that this algorithm finds a stationary point efficiently under mild conditions. The efficiency of the method is predicated on a dimensionality reduction enabled by a dual reformulation. Numerical experiments indicate that our approach has better computational and statistical properties than the state-of-the-art methods.

We develop an iterative (greedy) deep learning (DL) algorithm which is able to transform an arbitrary probability distribution function (PDF) into the target PDF. The model is based on iterative Optimal Transport of a series of 1D slices, matching on each slice the marginal PDF to the target. The axes of the orthogonal slices are chosen to maximize the PDF difference using Wasserstein distance at each iteration, which enables the algorithm to scale well to high dimensions. As special cases of this algorithm, we introduce two sliced iterative Normalizing Flow (SINF) models, which map from the data to the latent space (GIS) and vice versa (SIG). We show that SIG is able to generate high quality samples of image datasets, which match the GAN benchmarks, while GIS obtains competitive results on density estimation tasks compared to the density trained NFs, and is more stable, faster, and achieves higher p(x) when trained on small training sets. SINF approach deviates significantly from the current DL paradigm, as it is greedy and does not use concepts such as mini-batching, stochastic gradient descent and gradient back-propagation through deep layers.

We consider the problem of training User Verification (UV) models in federated setup, where each user has access to the data of only one class and user embeddings cannot be shared with the server or other users. To address this problem, we propose Federated User Verification (FedUV), a framework in which users jointly learn a set of vectors and maximize the correlation of their instance embeddings with a secret linear combination of those vectors. We show that choosing the linear combinations from the codewords of an error-correcting code allows users to collaboratively train the model without revealing their embedding vectors. We present the experimental results for user verification with voice, face, and handwriting data and show that FedUV is on par with existing approaches, while not sharing the embeddings with other users or the server.

\emph{Contrastive learning}~(CL) pretrains models in a pairwise manner, where given a data point, other data points are all regarded as dissimilar, including some that are \emph{semantically} similar. The issue has been addressed by properly weighting similar and dissimilar pairs as in \emph{positive-unlabeled learning}, so that the objective of CL is \emph{unbiased} and CL is \emph{consistent}. However, in this paper, we argue that this great solution is still not enough: its weighted objective \emph{hides} the issue where the semantically similar pairs are still pushed away; as CL is pretraining, this phenomenon is not our desideratum and might affect downstream tasks. To this end, we propose \emph{large-margin contrastive learning}~(LMCL) with \emph{distance polarization regularizer}, motivated by the distribution characteristic of pairwise distances in \emph{metric learning}. In LMCL, we can distinguish between \emph{intra-cluster} and \emph{inter-cluster} pairs, and then only push away inter-cluster pairs, which \emph{solves} the above issue explicitly. Theoretically, we prove a tighter error bound for LMCL; empirically, the superiority of LMCL is demonstrated across multiple domains, \emph{i.e.}, image classification, sentence representation, and reinforcement learning.

We study function approximation for episodic reinforcement learning with entropic risk measure. We first propose an algorithm with linear function approximation. Compared to existing algorithms, which suffer from improper regularization and regression biases, this algorithm features debiasing transformations in backward induction and regression procedures. We further propose an algorithm with general function approximation, which features implicit debiasing transformations. We prove that both algorithms achieve a sublinear regret and demonstrate a trade-off between generality and efficiency. Our analysis provides a unified framework for function approximation in risk-sensitive reinforcement learning, which leads to the first sublinear regret bounds in the setting.

Data sampling acts as a pivotal role in training deep learning models. However, an effective sampling schedule is difficult to learn due to its inherent high-dimension as a hyper-parameter. In this paper, we propose an AutoSampling method to automatically learn sampling schedules for model training, which consists of the multi-exploitation step aiming for optimal local sampling schedules and the exploration step for the ideal sampling distribution. More specifically, we achieve sampling schedule search with shortened exploitation cycle to provide enough supervision. In addition, we periodically estimate the sampling distribution from the learned sampling schedules and perturb it to search in the distribution space. The combination of two searches allows us to learn a robust sampling schedule. We apply our AutoSampling method to a variety of image classification tasks illustrating the effectiveness of the proposed method.

Credit assignment in reinforcement learning is the problem of measuring an action’s influence on future rewards. In particular, this requires separating skill from luck, i.e. disentangling the effect of an action on rewards from that of external factors and subsequent actions. To achieve this, we adapt the notion of counterfactuals from causality theory to a model-free RL setup. The key idea is to condition value functions on future events, by learning to extract relevant information from a trajectory. We formulate a family of policy gradient algorithms that use these future-conditional value functions as baselines or critics, and show that they are provably low variance. To avoid the potential bias from conditioning on future information, we constrain the hindsight information to not contain information about the agent's actions. We demonstrate the efficacy and validity of our algorithm on a number of illustrative and challenging problems.

Recently, a new recurrent neural network (RNN) named the Legendre Memory Unit (LMU) was proposed and shown to achieve state-of-the-art performance on several benchmark datasets. Here we leverage the linear time-invariant (LTI) memory component of the LMU to construct a simplified variant that can be parallelized during training (and yet executed as an RNN during inference), resulting in up to 200 times faster training. We note that our efficient parallelizing scheme is general and is applicable to any deep network whose recurrent components are linear dynamical systems. We demonstrate the improved accuracy of our new architecture compared to the original LMU and a variety of published LSTM and transformer networks across seven benchmarks. For instance, our LMU sets a new state-of-the-art result on psMNIST, and uses half the parameters while outperforming DistilBERT and LSTM models on IMDB sentiment analysis.

In many coordination problems, independently reasoning humans are able to discover mutually compatible policies. In contrast, independently trained self-play policies are often mutually incompatible. Zero-shot coordination (ZSC) has recently been proposed as a new frontier in multi-agent reinforcement learning to address this fundamental issue. Prior work approaches the ZSC problem by assuming players can agree on a shared learning algorithm but not on labels for actions and observations, and proposes other-play as an optimal solution. However, until now, this “label-free” problem has only been informally defined. We formalize this setting as the label-free coordination (LFC) problem by defining the label-free coordination game. We show that other-play is not an optimal solution to the LFC problem as it fails to consistently break ties between incompatible maximizers of the other-play objective. We introduce an extension of the algorithm, other-play with tie-breaking, and prove that it is optimal in the LFC problem and an equilibrium in the LFC game. Since arbitrary tie-breaking is precisely what the ZSC setting aims to prevent, we conclude that the LFC problem does not reflect the aims of ZSC. To address this, we introduce an alternative informal operationalization of ZSC as a starting point for future work.

As humans interact with autonomous agents to perform increasingly complicated, potentially risky tasks, it is important to be able to efficiently evaluate an agent's performance and correctness. In this paper we formalize and theoretically analyze the problem of efficient value alignment verification: how to efficiently test whether the behavior of another agent is aligned with a human's values? The goal is to construct a kind of "driver's test" that a human can give to any agent which will verify value alignment via a minimal number of queries. We study alignment verification problems with both idealized humans that have an explicit reward function as well as problems where they have implicit values. We analyze verification of exact value alignment for rational agents, propose and test heuristics for value alignment verification in gridworlds and a continuous autonomous driving domain, and prove that there exist sufficient conditions such that we can verify epsilon-alignment in any environment via a constant-query-complexity alignment test.

In this paper, we present a novel generative adversarial network (GAN) that can describe Markovian temporal dynamics. To generate stochastic sequential data, we introduce a novel stochastic differential equation-based conditional generator and spatial-temporal constrained discriminator networks. To stabilize the learning dynamics of the min-max type of the GAN objective function, we propose well-posed constraint terms for both networks. We also propose a novel conditional Markov Wasserstein distance to induce a pathwise Wasserstein distance. The experimental results demonstrate that our method outperforms state-of-the-art methods using several different types of data.

Federated Learning is a distributed learning setting where the goal is to train a centralized model with training data distributed over a large number of heterogeneous clients, each with unreliable and relatively slow network connections. A common optimization approach used in federated learning is based on the idea of local SGD: each client runs some number of SGD steps locally and then the updated local models are averaged to form the updated global model on the coordinating server. In this paper, we investigate the performance of an asynchronous version of local SGD wherein the clients can communicate with the server at arbitrary time intervals. Our main result shows that for smooth strongly convex and smooth nonconvex functions we achieve convergence rates that match the synchronous version that requires all clients to communicate simultaneously.

This paper introduces a new model to learn graph neural networks equivariant to rotations, translations, reflections and permutations called E(n)-Equivariant Graph Neural Networks (EGNNs). In contrast with existing methods, our work does not require computationally expensive higher-order representations in intermediate layers while it still achieves competitive or better performance. In addition, whereas existing methods are limited to equivariance on 3 dimensional spaces, our model is easily scaled to higher-dimensional spaces. We demonstrate the effectiveness of our method on dynamical systems modelling, representation learning in graph autoencoders and predicting molecular properties.

We propose a Multiscale Invertible Generative Network (MsIGN) and associated training algorithm that leverages multiscale structure to solve high-dimensional Bayesian inference. To address the curse of dimensionality, MsIGN exploits the low-dimensional nature of the posterior, and generates samples from coarse to fine scale (low to high dimension) by iteratively upsampling and refining samples. MsIGN is trained in a multi-stage manner to minimize the Jeffreys divergence, which avoids mode dropping in high-dimensional cases. On two high-dimensional Bayesian inverse problems, we show superior performance of MsIGN over previous approaches in posterior approximation and multiple mode capture. On the natural image synthesis task, MsIGN achieves superior performance in bits-per-dimension over baseline models and yields great interpret-ability of its neurons in intermediate layers.

Reconstructing continuous surfaces from 3D point clouds is a fundamental operation in 3D geometry processing. Several recent state-of-the-art methods address this problem using neural networks to learn signed distance functions (SDFs). In this paper, we introduce Neural-Pull, a new approach that is simple and leads to high quality SDFs. Specifically, we train a neural network to pull query 3D locations to their closest points on the surface using the predicted signed distance values and the gradient at the query locations, both of which are computed by the network itself. The pulling operation moves each query location with a stride given by the distance predicted by the network. Based on the sign of the distance, this may move the query location along or against the direction of the gradient of the SDF. This is a differentiable operation that allows us to update the signed distance value and the gradient simultaneously during training. Our outperforming results under widely used benchmarks demonstrate that we can learn SDFs more accurately and flexibly for surface reconstruction and single image reconstruction than the state-of-the-art methods. Our code and data are available at https://github.com/mabaorui/NeuralPull.

Many existing neural architecture search (NAS) solutions rely on downstream training for architecture evaluation, which takes enormous computations. Considering that these computations bring a large carbon footprint, this paper aims to explore a green (namely environmental-friendly) NAS solution that evaluates architectures without training. Intuitively, gradients, induced by the architecture itself, directly decide the convergence and generalization results. It motivates us to propose the gradient kernel hypothesis: Gradients can be used as a coarse-grained proxy of downstream training to evaluate random-initialized networks. To support the hypothesis, we conduct a theoretical analysis and find a practical gradient kernel that has good correlations with training loss and validation performance. According to this hypothesis, we propose a new kernel based architecture search approach KNAS. Experiments show that KNAS achieves competitive results with orders of magnitude faster than ``train-then-test'' paradigms on image classification tasks. Furthermore, the extremely low search cost enables its wide applications. The searched network also outperforms strong baseline RoBERTA-large on two text classification tasks.

Although model-based and model-free approaches to learning the control of systems have achieved impressive results on standard benchmarks, generalization to task variations is still lacking. Recent results suggest that generalization for standard architectures improves only after obtaining exhaustive amounts of data. We give evidence that generalization capabilities are in many cases bottlenecked by the inability to generalize on the combinatorial aspects of the problem. We show that, for a certain subclass of the MDP framework, this can be alleviated by a neuro-algorithmic policy architecture that embeds a time-dependent shortest path solver in a deep neural network. Trained end-to-end via blackbox-differentiation, this method leads to considerable improvement in generalization capabilities in the low-data regime.

We define deep kernel processes in which positive definite Gram matrices are progressively transformed by nonlinear kernel functions and by sampling from (inverse) Wishart distributions. Remarkably, we find that deep Gaussian processes (DGPs), Bayesian neural networks (BNNs), infinite BNNs, and infinite BNNs with bottlenecks can all be written as deep kernel processes. For DGPs the equivalence arises because the Gram matrix formed by the inner product of features is Wishart distributed, and as we show, standard isotropic kernels can be written entirely in terms of this Gram matrix --- we do not need knowledge of the underlying features. We define a tractable deep kernel process, the deep inverse Wishart process, and give a doubly-stochastic inducing-point variational inference scheme that operates on the Gram matrices, not on the features, as in DGPs. We show that the deep inverse Wishart process gives superior performance to DGPs and infinite BNNs on fully-connected baselines.

A key challenge on the path to developing agents that learn complex human-like behavior is the need to quickly and accurately quantify human-likeness. While human assessments of such behavior can be highly accurate, speed and scalability are limited. We address these limitations through a novel automated Navigation Turing Test (ANTT) that learns to predict human judgments of human-likeness. We demonstrate the effectiveness of our automated NTT on a navigation task in a complex 3D environment. We investigate six classification models to shed light on the types of architectures best suited to this task, and validate them against data collected through a human NTT. Our best models achieve high accuracy when distinguishing true human and agent behavior. At the same time, we show that predicting finer-grained human assessment of agents’ progress towards human-like behavior remains unsolved. Our work takes an important step towards agents that more effectively learn complex human-like behavior.

With a better understanding of the loss surfaces for multilayer networks, we can build more robust and accurate training procedures. Recently it was discovered that independently trained SGD solutions can be connected along one-dimensional paths of near-constant training loss. In this paper, we in fact demonstrate the existence of mode-connecting simplicial complexes that form multi-dimensional manifolds of low loss, connecting many independently trained models. Building on this discovery, we show how to efficiently construct simplicial complexes for fast ensembling, outperforming independently trained deep ensembles in accuracy, calibration, and robustness to dataset shift. Notably, our approach is easy to apply and only requires a few training epochs to discover a low-loss simplex.

The time and effort involved in hand-designing deep neural networks is immense. This has prompted the development of Neural Architecture Search (NAS) techniques to automate this design. However, NAS algorithms tend to be slow and expensive; they need to train vast numbers of candidate networks to inform the search process. This could be alleviated if we could partially predict a network's trained accuracy from its initial state. In this work, we examine the overlap of activations between datapoints in untrained networks and motivate how this can give a measure which is usefully indicative of a network’s trained performance. We incorporate this measure into a simple algorithm that allows us to search for powerful networks without any training in a matter of seconds on a single GPU, and verify its effectiveness on NAS-Bench-101, NAS-Bench-201, NATS-Bench, and Network Design Spaces. Our approach can be readily combined with more expensive search methods; we examine a simple adaptation of regularised evolutionary search. Code for reproducing our experiments is available at https://github.com/BayesWatch/nas-without-training.

Convolutional architectures have proven extremely successful for vision tasks. Their hard inductive biases enable sample-efficient learning, but come at the cost of a potentially lower performance ceiling. Vision Transformers (ViTs) rely on more flexible self-attention layers, and have recently outperformed CNNs for image classification. However, they require costly pre-training on large external datasets or distillation from pre-trained convolutional networks. In this paper, we ask the following question: is it possible to combine the strengths of these two architectures while avoiding their respective limitations? To this end, we introduce gated positional self-attention (GPSA), a form of positional self-attention which can be equipped with a ``soft" convolutional inductive bias. We initialise the GPSA layers to mimic the locality of convolutional layers, then give each attention head the freedom to escape locality by adjusting a gating parameter regulating the attention paid to position versus content information. The resulting convolutional-like ViT architecture, ConViT, outperforms the DeiT on ImageNet, while offering a much improved sample efficiency. We further investigate the role of locality in learning by first quantifying how it is encouraged in vanilla self-attention layers, then analysing how it is escaped in GPSA layers. We conclude by presenting various ablations to better understand the success of the ConViT. Our code and models are released publicly at https://github.com/facebookresearch/convit.

Off-policy reinforcement learning (RL) holds the promise of sample-efficient learning of decision-making policies by leveraging past experience. However, in the offline RL setting -- where a fixed collection of interactions are provided and no further interactions are allowed -- it has been shown that standard off-policy RL methods can significantly underperform. In this work, we closely investigate an important simplification of BCQ (Fujimoto et al., 2018) -- a prior approach for offline RL -- removing a heuristic design choice. Importantly, in contrast to their original theoretical considerations, we derive this simplified algorithm through the introduction of a novel backup operator, Expected-Max Q-Learning (EMaQ), which is more closely related to the resulting practical algorithm. Specifically, in addition to the distribution support, EMaQ explicitly considers the number of samples and the proposal distribution, allowing us to derive new sub-optimality bounds. In the offline RL setting -- the main focus of this work -- EMaQ matches and outperforms prior state-of-the-art in the D4RL benchmarks (Fu et al., 2020). In the online RL setting, we demonstrate that EMaQ is competitive with Soft Actor Critic (SAC). The key contributions of our empirical findings are demonstrating the importance of careful generative model design for estimating behavior policies, and an intuitive notion of complexity for offline RL problems. With its simple interpretation and fewer moving parts, such as no explicit function approximator representing the policy, EMaQ serves as a strong yet easy to implement baseline for future work.

Deep Metric Learning (DML) provides a crucial tool for visual similarity and zero-shot retrieval applications by learning generalizing embedding spaces, although recent work in DML has shown strong performance saturation across training objectives. However, generalization capacity is known to scale with the embedding space dimensionality. Unfortunately, high dimensional embeddings also create higher retrieval cost for downstream applications. To remedy this, we propose S2SD - Simultaneous Similarity-based Self-distillation. S2SD extends DML with knowledge distillation from auxiliary, high-dimensional embedding and feature spaces to leverage complementary context during training while retaining test-time cost and with negligible changes to the training time. Experiments and ablations across different objectives and standard benchmarks show S2SD offering highly significant improvements of up to 7% in Recall@1, while also setting a new state-of-the-art.

A fundamental challenge in multiagent reinforcement learning is to learn beneficial behaviors in a shared environment with other simultaneously learning agents. In particular, each agent perceives the environment as effectively non-stationary due to the changing policies of other agents. Moreover, each agent is itself constantly learning, leading to natural non-stationarity in the distribution of experiences encountered. In this paper, we propose a novel meta-multiagent policy gradient theorem that directly accounts for the non-stationary policy dynamics inherent to multiagent learning settings. This is achieved by modeling our gradient updates to consider both an agent’s own non-stationary policy dynamics and the non-stationary policy dynamics of other agents in the environment. We show that our theoretically grounded approach provides a general solution to the multiagent learning problem, which inherently comprises all key aspects of previous state of the art approaches on this topic. We test our method on a diverse suite of multiagent benchmarks and demonstrate a more efficient ability to adapt to new agents as they learn than baseline methods across the full spectrum of mixed incentive, competitive, and cooperative domains.

Biological systems understand the world by simultaneously processing high-dimensional inputs from modalities as diverse as vision, audition, touch, proprioception, etc. The perception models used in deep learning on the other hand are designed for individual modalities, often relying on domain-specific assumptions such as the local grid structures exploited by virtually all existing vision models. These priors introduce helpful inductive biases, but also lock models to individual modalities. In this paper we introduce the Perceiver – a model that builds upon Transformers and hence makes few architectural assumptions about the relationship between its inputs, but that also scales to hundreds of thousands of inputs, like ConvNets. The model leverages an asymmetric attention mechanism to iteratively distill inputs into a tight latent bottleneck, allowing it to scale to handle very large inputs. We show that this architecture is competitive with or outperforms strong, specialized models on classification tasks across various modalities: images, point clouds, audio, video and video+audio. The Perceiver obtains performance comparable to ResNet-50 and ViT on ImageNet without 2D convolutions by directly attending to 50,000 pixels. It is also competitive in all modalities in AudioSet.

We propose a novel approximation hierarchy for cardinality-constrained, convex quadratic programs that exploits the rank-dominating eigenvectors of the quadratic matrix. Each level of approximation admits a min-max characterization whose objective function can be optimized over the binary variables analytically, while preserving convexity in the continuous variables. Exploiting this property, we propose two scalable optimization algorithms, coined as the best response" and thedual program", that can efficiently screen the potential indices of the nonzero elements of the original program. We show that the proposed methods are competitive with the existing screening methods in the current sparse regression literature, and it is particularly fast on instances with high number of measurements in experiments with both synthetic and real datasets.

Fragment libraries are often used in protein structure prediction, simulation and design as a means to significantly reduce the vast conformational search space. Current state-of-the-art methods for fragment library generation do not properly account for aleatory and epistemic uncertainty, respectively due to the dynamic nature of proteins and experimental errors in protein structures. Additionally, they typically rely on information that is not generally or readily available, such as homologous sequences, related protein structures and other complementary information. To address these issues, we developed BIFROST, a novel take on the fragment library problem based on a Deep Markov Model architecture combined with directional statistics for angular degrees of freedom, implemented in the deep probabilistic programming language Pyro. BIFROST is a probabilistic, generative model of the protein backbone dihedral angles conditioned solely on the amino acid sequence. BIFROST generates fragment libraries with a quality on par with current state-of-the-art methods at a fraction of the run-time, while requiring considerably less information and allowing efficient evaluation of probabilities.

Can a neural network minimizing cross-entropy learn linearly separable data? Despite progress in the theory of deep learning, this question remains unsolved. Here we prove that SGD globally optimizes this learning problem for a two-layer network with Leaky ReLU activations. The learned network can in principle be very complex. However, empirical evidence suggests that it often turns out to be approximately linear. We provide theoretical support for this phenomenon by proving that if network weights converge to two weight clusters, this will imply an approximately linear decision boundary. Finally, we show a condition on the optimization that leads to weight clustering. We provide empirical results that validate our theoretical analysis.

One of the ways that machine learning algorithms can help control the spread of an infectious disease is by building models that predict who is likely to become infected making them good candidates for preemptive interventions. In this work we ask: can we build reliable infection prediction models when the observed data is collected under limited, and biased testing that prioritizes testing symptomatic individuals? Our analysis suggests that when the infection is highly transmissible, incomplete testing might be sufficient to achieve good out-of-sample prediction error. Guided by this insight, we develop an algorithm that predicts infections, and show that it outperforms baselines on simulated data. We apply our model to data from a large hospital to predict Clostridioides difficile infections; a communicable disease that is characterized by both symptomatically infected and asymptomatic (i.e., untested) carriers. Using a proxy instead of the unobserved untested-infected state, we show that our model outperforms benchmarks in predicting infections.

Many important real-world problems have action spaces that are high-dimensional, continuous or both, making full enumeration of all possible actions infeasible. Instead, only small subsets of actions can be sampled for the purpose of policy evaluation and improvement. In this paper, we propose a general framework to reason in a principled way about policy evaluation and improvement over such sampled action subsets. This sample-based policy iteration framework can in principle be applied to any reinforcement learning algorithm based upon policy iteration. Concretely, we propose Sampled MuZero, an extension of the MuZero algorithm that is able to learn in domains with arbitrarily complex action spaces by planning over sampled actions. We demonstrate this approach on the classical board game of Go and on two continuous control benchmark domains: DeepMind Control Suite and Real-World RL Suite.

Various gradient compression schemes have been proposed to mitigate the communication cost in distributed training of large scale machine learning models. Sign-based methods, such as signSGD (Bernstein et al., 2018), have recently been gaining popularity because of their simple compression rule and connection to adaptive gradient methods, like ADAM. In this paper, we analyze sign-based methods for non-convex optimization in three key settings: (i) standard single node, (ii) parallel with shared data and (iii) distributed with partitioned data. For single machine case, we generalize the previous analysis of signSGD relying on intuitive bounds on success probabilities and allowing even biased estimators. Furthermore, we extend the analysis to parallel setting within a parameter server framework, where exponentially fast noise reduction is guaranteed with respect to number of nodes, maintaining $1$-bit compression in both directions and using small mini-batch sizes. Next, we identify a fundamental issue with signSGD to converge in distributed environment. To resolve this issue, we propose a new sign-based method, {\em Stochastic Sign Descent with Momentum (SSDM)}, which converges under standard bounded variance assumption with the optimal asymptotic rate. We validate several aspects of our theoretical findings with numerical experiments.

Imitation learning enables agents to reuse and adapt the hard-won expertise of others, offering a solution to several key challenges in learning behavior. Although it is easy to observe behavior in the real-world, the underlying actions may not be accessible. We present a new method for imitation solely from observations that achieves comparable performance to experts on challenging continuous control tasks while also exhibiting robustness in the presence of observations unrelated to the task. Our method, which we call FORM (for "Future Observation Reward Model") is derived from an inverse RL objective and imitates using a model of expert behavior learned by generative modelling of the expert's observations, without needing ground truth actions. We show that FORM performs comparably to a strong baseline IRL method (GAIL) on the DeepMind Control Suite benchmark, while outperforming GAIL in the presence of task-irrelevant features.

We consider the problem of molecular graph generation using deep models. While graphs are discrete, most existing methods use continuous latent variables, resulting in inaccurate modeling of discrete graph structures. In this work, we propose GraphDF, a novel discrete latent variable model for molecular graph generation based on normalizing flow methods. GraphDF uses invertible modulo shift transforms to map discrete latent variables to graph nodes and edges. We show that the use of discrete latent variables reduces computational costs and eliminates the negative effect of dequantization. Comprehensive experimental results show that GraphDF outperforms prior methods on random generation, property optimization, and constrained optimization tasks.

Offline Reinforcement Learning methods seek to learn a policy from logged transitions of an environment, without any interaction. In the presence of function approximation, and under the assumption of limited coverage of the state-action space of the environment, it is necessary to enforce the policy to visit state-action pairs close to the support of logged transitions. In this work, we propose an iterative procedure to learn a pseudometric (closely related to bisimulation metrics) from logged transitions, and use it to define this notion of closeness. We show its convergence and extend it to the function approximation setting. We then use this pseudometric to define a new lookup based bonus in an actor-critic algorithm: PLOFF. This bonus encourages the actor to stay close, in terms of the defined pseudometric, to the support of logged transitions. Finally, we evaluate the method on hand manipulation and locomotion tasks.

Recent advances in Transformer models allow for unprecedented sequence lengths, due to linear space and time complexity. In the meantime, relative positional encoding (RPE) was proposed as beneficial for classical Transformers and consists in exploiting lags instead of absolute positions for inference. Still, RPE is not available for the recent linear-variants of the Transformer, because it requires the explicit computation of the attention matrix, which is precisely what is avoided by such methods. In this paper, we bridge this gap and present Stochastic Positional Encoding as a way to generate PE that can be used as a replacement to the classical additive (sinusoidal) PE and provably behaves like RPE. The main theoretical contribution is to make a connection between positional encoding and cross-covariance structures of correlated Gaussian processes. We illustrate the performance of our approach on the Long-Range Arena benchmark and on music generation.

Likelihood is a standard estimate for outlier detection. The specific role of the normalization constraint is to ensure that the out-of-distribution (OOD) regime has a small likelihood when samples are learned using maximum likelihood. Because autoencoders do not possess such a process of normalization, they often fail to recognize outliers even when they are obviously OOD. We propose the Normalized Autoencoder (NAE), a normalized probabilistic model constructed from an autoencoder. The probability density of NAE is defined using the reconstruction error of an autoencoder, which is differently defined in the conventional energy-based model. In our model, normalization is enforced by suppressing the reconstruction of negative samples, significantly improving the outlier detection performance. Our experimental results confirm the efficacy of NAE, both in detecting outliers and in generating in-distribution samples.

Low-dimensional node embeddings play a key role in analyzing graph datasets. However, little work studies exactly what information is encoded by popular embedding methods, and how this information correlates with performance in downstream learning tasks. We tackle this question by studying whether embeddings can be inverted to (approximately) recover the graph used to generate them. Focusing on a variant of the popular DeepWalk method \cite{PerozziAl-RfouSkiena:2014, QiuDongMa:2018}, we present algorithms for accurate embedding inversion -- i.e., from the low-dimensional embedding of a graph $G$, we can find a graph $\tilde G$ with a very similar embedding. We perform numerous experiments on real-world networks, observing that significant information about $G$, such as specific edges and bulk properties like triangle density, is often lost in $\tilde G$. However, community structure is often preserved or even enhanced. Our findings are a step towards a more rigorous understanding of exactly what information embeddings encode about the input graph, and why this information is useful for learning tasks.

The performance of reinforcement learning depends upon designing an appropriate action space, where the effect of each action is measurable, yet, granular enough to permit flexible behavior. So far, this process involved non-trivial user choices in terms of the available actions and their execution frequency. We propose a novel framework for reinforcement learning that effectively lifts such constraints. Within our framework, agents learn effective behavior over a routine space: a new, higher-level action space, where each routine represents a set of 'equivalent' sequences of granular actions with arbitrary length. Our routine space is learned end-to-end to facilitate the accomplishment of underlying off-policy reinforcement learning objectives. We apply our framework to two state-of-the-art off-policy algorithms and show that the resulting agents obtain relevant performance improvements while requiring fewer interactions with the environment per episode, improving computational efficiency.

We show how fitting sparse linear models over learned deep feature representations can lead to more debuggable neural networks. These networks remain highly accurate while also being more amenable to human interpretation, as we demonstrate quantitatively and via human experiments. We further illustrate how the resulting sparse explanations can help to identify spurious correlations, explain misclassifications, and diagnose model biases in vision and language tasks.

Transfer in reinforcement learning is usually achieved through generalisation across tasks. Whilst many studies have investigated transferring knowledge when the reward function changes, they have assumed that the dynamics of the environments remain consistent. Many real-world RL problems require transfer among environments with different dynamics. To address this problem, we propose an approach based on successor features in which we model successor feature functions with Gaussian Processes permitting the source successor features to be treated as noisy measurements of the target successor feature function. Our theoretical analysis proves the convergence of this approach as well as the bounded error on modelling successor feature functions with Gaussian Processes in environments with both different dynamics and rewards. We demonstrate our method on benchmark datasets and show that it outperforms current baselines.

Recently, local SGD has got much attention and been extensively studied in the distributed learning community to overcome the communication bottleneck problem. However, the superiority of local SGD to minibatch SGD only holds in quite limited situations. In this paper, we study a new local algorithm called Bias-Variance Reduced Local SGD (BVR-L-SGD) for nonconvex distributed optimization. Algorithmically, our proposed bias and variance reduced local gradient estimator fully utilizes small second-order heterogeneity of local objectives and suggests randomly picking up one of the local models instead of taking the average of them when workers are synchronized. Theoretically, under small heterogeneity of local objectives, we show that BVR-L-SGD achieves better communication complexity than both the previous non-local and local methods under mild conditions, and particularly BVR-L-SGD is the first method that breaks the barrier of communication complexity $\Theta(1/\varepsilon)$ for general nonconvex smooth objectives when the heterogeneity is small and the local computation budget is large. Numerical results are given to verify the theoretical findings and give empirical evidence of the superiority of our method.

Spiking Neural Network (SNN) has been recognized as one of the next generation of neural networks. Conventionally, SNN can be converted from a pre-trained ANN by only replacing the ReLU activation to spike activation while keeping the parameters intact. Perhaps surprisingly, in this work we show that a proper way to calibrate the parameters during the conversion of ANN to SNN can bring significant improvements. We introduce SNN Calibration, a cheap but extraordinarily effective method by leveraging the knowledge within a pre-trained Artificial Neural Network (ANN). Starting by analyzing the conversion error and its propagation through layers theoretically, we propose the calibration algorithm that can correct the error layer-by-layer. The calibration only takes a handful number of training data and several minutes to finish. Moreover, our calibration algorithm can produce SNN with state-of-the-art architecture on the large-scale ImageNet dataset, including MobileNet and RegNet. Extensive experiments demonstrate the effectiveness and efficiency of our algorithm. For example, our advanced pipeline can increase up to 69% top-1 accuracy when converting MobileNet on ImageNet compared to baselines. Codes are released at https://github.com/yhhhli/SNN_Calibration.

We study a class of convex-concave saddle-point problems of the form $\min_x\max_y \langle Kx,y\rangle+f_{\cal P}(x)-h^*(y)$ where $K$ is a linear operator, $f_{\cal P}$ is the sum of a convex function $f$ with a Lipschitz-continuous gradient and the indicator function of a bounded convex polytope ${\cal P}$, and $h^\ast$ is a convex (possibly nonsmooth) function. Such problem arises, for example, as a Lagrangian relaxation of various discrete optimization problems. Our main assumptions are the existence of an efficient {\em linear minimization oracle} ($lmo$) for $f_{\cal P}$ and an efficient {\em proximal map} ($prox$) for $h^*$ which motivate the solution via a blend of proximal primal-dual algorithms and Frank-Wolfe algorithms. In case $h^*$ is the indicator function of a linear constraint and function $f$ is quadratic, we show a $O(1/n^2)$ convergence rate on the dual objective, requiring $O(n \log n)$ calls of $lmo$. If the problem comes from the constrained optimization problem $\min_{x\in\mathbb R^d}\{f_{\cal P}(x)\:|\:Ax-b=0\}$ then we additionally get bound $O(1/n^2)$ both on the primal gap and on the infeasibility gap. In the most general case, we show a $O(1/n)$ convergence rate of the primal-dual gap again requiring $O(n\log n)$ calls of $lmo$. To the best of our knowledge, this improves on the known convergence rates for the considered class of saddle-point problems. We show applications to labeling problems frequently appearing in machine learning and computer vision.

The dot product self-attention is known to be central and indispensable to state-of-the-art Transformer models. But is it really required? This paper investigates the true importance and contribution of the dot product-based self-attention mechanism on the performance of Transformer models. Via extensive experiments, we find that (1) random alignment matrices surprisingly perform quite competitively and (2) learning attention weights from token-token (query-key) interactions is useful but not that important after all. To this end, we propose \textsc{Synthesizer}, a model that learns synthetic attention weights without token-token interactions. In our experiments, we first show that simple Synthesizers achieve highly competitive performance when compared against vanilla Transformer models across a range of tasks, including machine translation, language modeling, text generation and GLUE/SuperGLUE benchmarks. When composed with dot product attention, we find that Synthesizers consistently outperform Transformers. Moreover, we conduct additional comparisons of Synthesizers against Dynamic Convolutions, showing that simple Random Synthesizer is not only $60\%$ faster but also improves perplexity by a relative $3.5\%$. Finally, we show that simple factorized Synthesizers can outperform Linformers on encoding only tasks.

VDN and QMIX are two popular value-based algorithms for cooperative MARL that learn a centralized action value function as a monotonic mixing of per-agent utilities. While this enables easy decentralization of the learned policy, the restricted joint action value function can prevent them from solving tasks that require significant coordination between agents at a given timestep. We show that this problem can be overcome by improving the joint exploration of all agents during training. Specifically, we propose a novel MARL approach called Universal Value Exploration (UneVEn) that learns a set of related tasks simultaneously with a linear decomposition of universal successor features. With the policies of already solved related tasks, the joint exploration process of all agents can be improved to help them achieve better coordination. Empirical results on a set of exploration games, challenging cooperative predator-prey tasks requiring significant coordination among agents, and StarCraft II micromanagement benchmarks show that UneVEn can solve tasks where other state-of-the-art MARL methods fail.

Continual learning—the ability to learn many tasks in sequence—is critical for artificial learning systems. Yet standard training methods for deep networks often suffer from catastrophic forgetting, where learning new tasks erases knowledge of the earlier tasks. While catastrophic forgetting labels the problem, the theoretical reasons for interference between tasks remain unclear. Here, we attempt to narrow this gap between theory and practice by studying continual learning in the teacher-student setup. We extend previous analytical work on two-layer networks in the teacher-student setup to multiple teachers. Using each teacher to represent a different task, we investigate how the relationship between teachers affects the amount of forgetting and transfer exhibited by the student when the task switches. In line with recent work, we find that when tasks depend on similar features, intermediate task similarity leads to greatest forgetting. However, feature similarity is only one way in which tasks may be related. The teacher-student approach allows us to disentangle task similarity at the level of \emph{readouts} (hidden-to-output weights) as well as \emph{features} (input-to-hidden weights). We find a complex interplay between both types of similarity, initial transfer/forgetting rates, maximum transfer/forgetting, and the long-time (post-switch) amount of transfer/forgetting. Together, these results help illuminate the diverse factors contributing to catastrophic forgetting.

We introduce Phasic Policy Gradient (PPG), a reinforcement learning framework which modifies traditional on-policy actor-critic methods by separating policy and value function training into distinct phases. In prior methods, one must choose between using a shared network or separate networks to represent the policy and value function. Using separate networks avoids interference between objectives, while using a shared network allows useful features to be shared. PPG is able to achieve the best of both worlds by splitting optimization into two phases, one that advances training and one that distills features. PPG also enables the value function to be more aggressively optimized with a higher level of sample reuse. Compared to PPO, we find that PPG significantly improves sample efficiency on the challenging Procgen Benchmark.