I'll talk about one specific problem I have with the field: scale. Many papers fix an architecture and try to improve log-likelihood, comparing to the original base architecture regardless of how much additional compute is used to outperform the original model. Yet, if we adjust for scale—for example, compare an ensemble of size 10 to a model scaled up 10x—we'd see improvements significantly diminish or vanish altogether. Ultimately, we should be examining the frontier of uncertainty-robustness performance as a function of compute. I'll substantiate this perspective with a few works with colleagues. These works advance the frontier with efficient ensembles alongside priors and inductive biases; and we'll examine uncertainty properties of existing giant models.

[ protected link dropped ] [ protected link dropped ]

OOD generalization is a very difficult problem. Instead of tackling it head on, this talk argues that, when considering the current strengths and weaknesses of deep learning, we should consider an alternative approach which tries to dodge the problem altogether. If we can develop scalable pre-training methods that can leverage large and highly varied data sources, there is a hope that many examples (which would have been OOD for standard ML datasets) will have at least some relevant training data, removing the need for elusive OOD capabilities.

[ protected link dropped ] [ protected link dropped ]

Machine learning models deployed in the real world constantly face distribution shifts, yet current models are not robust to these shifts; they can perform well when the train and test distributions are identical, but still have their performance plummet when evaluated on a different test distribution. In this talk, I will discuss methods and benchmarks for improving robustness to distribution shifts. First, we consider the problem of spurious correlations and show how to mitigate it with a combination of distributionally robust optimization (DRO) and controlling model complexity---e.g., through strong L2 regularization, early stopping, or underparameterization. Second, we present WILDS, a curated and diverse collection of 10 datasets with real-world distribution shifts, that aims to address the under-representation of real-world shifts in the datasets widely used in the ML community today. We observe that existing methods fail to mitigate performance drops due to these distribution shifts, underscoring the need for new training methods that produce models which are more robust to the types of distribution shifts that arise in practice.

Aggregate evaluations of deep learning models on popular benchmarks have incentivized the creation of bigger models that are more accurate on iid data. As the research community is realizing that these models do not generalize out of distribution, the trend has shifted to evaluations on adversarially constructed, unnatural datasets. However, both these extremities have limitations when it comes to meeting the goals of evaluation. In this talk, I propose that the goal of evaluation is to inform next action to a user in the form of 1) further analysis or 2) model patching. Thinking of evaluation as an iterative process dovetails with these goals. Our work on Robustness Gym (RG) proposes an iterative process of evaluation and explains how that enables a user to iterate on their model development process. I will give two concrete examples in NLP demonstrating how RG supports the aforementioned evaluation goals. Towards the end of the talk, I will discuss some caveats associated with evaluating pre-trained language models (PLMs) and in particular focus on the problem of input contamination, giving examples from our work on SummVis. Using these examples from RG and SummVis, I hope to draw attention to the limitations of current evaluations and the need for a more thorough process that helps us gain a better understanding of our deep learning models.

Novelty detection, i.e., identifying whether a given sample is drawn from outside the training distribution, is essential for reliable machine learning. To this end, there have been many attempts at learning a representation well-suited for novelty detection and designing a score based on such representation. In this talk, I will present a simple, yet effective method named contrasting shifted instances (CSI), inspired by the recent success on contrastive learning of visual representations. Specifically, in addition to contrasting a given sample with other instances as in conventional contrastive learning methods, our training scheme contrasts the sample with distributionally-shifted augmentations of itself. Based on this, we propose a new detection score that is specific to the proposed training scheme. Our experiments demonstrate the superiority of our method under various novelty detection scenarios, including unlabeled one-class, unlabeled multi-class and labeled multi-class settings, with various image benchmark datasets. This is a joint work with Jihoon Tack, Sangwoo Mo and Jongheon Jeong (all from KAIST).