As the electricity grid transitions to using an increasing amount of renewable energy, aspects of its operation will change in order to manage variability in spatio-temporal mismatches in supply and demand. These mismatches can lead to power grid congestion due to limitations in flows between areas. Numerous approaches for dealing with this are either already operational or are actively being researched. These solutions take a variety of forms such as physical technologies (batteries, dynamic power line ratings), algorithms (optimal power flow solvers) and electricity market design (virtual bidding, ancillary services).
Machine Learning can play a role in all of these. Our focus is on electricity markets and using ML to forecast the divergence between Day-Ahead and Real-Time nodal prices. This helps to close the gap between expected conditions used for planning and those actually experienced, leading to lower emissions and increased reliability of the power grid. As other approaches are also trying to minimize this divergence, this becomes a task of forecasting the forecast error of other market participants, or at least their limitations in mitigating diverging conditions. This is further complicated by feedback loops between approaches acting at different system levels and the continuing evolution of the grid's underlying behaviour and control mechanisms.
In addressing these challenges, we employ a spectrum of approaches, from physics-aware models including electrical constraints of the grid, to physics-agnostic models which help identify the inherent uncertainty and biases in planning for grid reliability.
AI development today rests on three pillars: algorithms, hardware, and data. Ironically, the further AI moves towards new application areas, the more it depends on human efforts: more and more often data for training and validating AI models cannot be collected in any other way than by humans.
AI solutions require data for training and validating models that are not only high-quality and scalable to support growing industry needs but also flexible enough to support a large variety of use cases and data collection scenarios.
Toloka's mission is to create an environment for AI data production that is fully aligned with industry needs: quality, scalability, flexibility.
As a result, Toloka is a multifaceted solution with:
- a global crowdforce of 9 million Tolokers with around 200,000 active on the platform every month
- multiple methods and mechanisms for advanced automated quality control at scale, available for any platform using the Crowd-Kit library for Python
- instruments for integrating the crowd into the ML production process using the Toloka-Kit library for Python
- academic research and education initiatives in the field of Crowd Science for ML specialists
The Toloka workshop aims to cover these aspects and provide a comprehensive picture of how crowdsourcing can be applied to real life AI production.
The theme of this workshop will be real world applications of reinforcement learning. We will give a demo/tutorial of the latest features and additions to Azure Personalizer, an award winning and easy to use RL cloud service (https://azure.microsoft.com/en-us/services/cognitive-services/personalizer/). The first release of Personalizer was presented in an ICML 2019 workshop. We will present the latest additions to the service, including multi-slot personalization.
We will demonstrate how to leverage the latest release of Vowpal Wabbit, an open source machine learning library (https://vowpalwabbit.org/). It provides fast, scalable machine learning and has unique capabilities such as learning to search, active learning, contextual memory, and extreme multiclass learning. It has a focus on reinforcement learning and provides production ready implementations of Contextual Bandit algorithms.
This part will include:
- Demo on COBA - benchmarking framework for CB algorithms (https://github.com/VowpalWabbit/coba)
- Using Panda Dataframes with pyVW
- AutoML & pyVW
- Hands on demo of continuous actions with VW
- Integrating VW with Apache Spark
The current state of the art in machine learning relies on past patterns and correlations to make predictions of the future. This approach can work in static environments and for closed problems with fixed rules. However, it does not for dynamic systems such as financial time-series. In order to make consistently accurate predictions about the future and to achieve true artificial intelligence, the development of new science that enables machines to understand cause and effect is required. Understanding true causal drivers enables causal AI to navigate complex and dynamic systems, being able to perform as its environment changes. In addition, causal AI is capable of ‘imagining’ scenarios it has not encountered in the past, allowing it to simulate counterfactual worlds to learn from, instead of relying solely on ‘training’ data. Perhaps most interestingly, understanding causality gives an AI the ability to interact with humans more deeply, being able to explain its ‘thought process’ and integrate human knowledge.
At causaLens we are building the world’s largest Causal AI research lab to accelerate progress in this powerful science. This talk will present the current challenges causal AI must overcome to unleash its full potential, as well as the latest progress made towards achieving those goals. Finally, some examples of the positive impact this science is already having in the field will be shared.
We here introduce unique opportunities for research and collaboration in AI Algorithms, Health, Traffic, and Weather at the Institute of Advanced Research in Artificial Intelligence (IARAI), followed by an Open Panel for discussion. At IARAI, we are building a unique environment of world class researchers and industrial-scale real-world data openly shared with the scientific community. In this session, we showcase (1) unique data sets from urban traffic which have featured in the NeurIPS competition track since 2019, (2) high-resolution multi-channel movies from weather satellites newly available now, and (3) plans for freely sharing data from over 10 million electronic health records with the academic community in a secure way. This is complemented by an outline of recent highlights in algorithm development, linking attention in Transformers with associative memory, achieving world-class cross-domain few-shot learning performance with a novel light-weight approach, and discussing new insights on mixing layers. We close by outlining future opportunities of engaging at the institute and introduce the latest Scientific Programmes, followed by a short panel discussion on the hottest open topics in the field.
Virtual Reality (VR) allows users to immerse themselves into imaginative worlds and enjoy appearances and behaviors that can be unlike anything in the real world. Users simply put on a Head Mounted Display (HMDs), pick up hand-held controllers, and enter a virtual environment where they can see themselves, many times represented by a virtual avatar, interact with content displayed around them using natural motion. This synchronized mapping of the user’s body and the self-avatar’s motions enhances embodiment (i.e., a sense of ownership of the virtual avatar), and the direct control of the motions develops a sense of agency (i.e., a feeling of control over the avatar’s actions and their consequences). Both are important characteristics of the immersions and captivating VR experiences.
However, while many authored media such as movies, games, and social applications has used character motion stylization to generate impressive character motions beyond the gamer/user/viewer's abilities, Virtual Reality are bereft of such motion accentuations of the self-avatars. The user may find herself in a novel imaginative environment where the law of physicals behaves differently, yet her avatar motion follows her own body motions which are limited by physical laws, her experience, her environment, social limitations, physical limitations and more.
I will present CoolMoves, a novel proof-of-concept system for expressive and accentuated full-body motion synthesis of a user’s virtual avatar in real-time., while maintaining high embodiment and agency. CoolMoves uses the limited sensing, such as done by current consumer-grade VR systems, specifically headset and hand positions, and synthesizes full- body motions. Motion trajectories for each joint are synthesized through our processing pipeline that takes an exist motion capture database as input and, in real time, loosely matches segments of the user’s motions and accentuates as well as blends them to animate the user’s avatar in VR. The system can extrapolate free-ranging motions that may do not originally appear in the motion data base, enabling the user to be fully expressive in her motion, yet maintain requested stylization.
I see Cool Moves as a major step in effort make VR a preferable work modality with potential to let to set their work with a personal fit, users can achieve more with less physical effort, and hopefully level the plain field for people with physical limitations.
Video technology has revolutionized how we create and consume media. Advancements in video compression, which provide enhanced video quality with less bits, have led to broad video adoption across a wide range of devices and services. In fact, it’s expected that 82% of Internet traffic will be video by 2022 . With this explosive growth of video traffic, video coding technology enhancements are crucial for providing entertainment, enhancing collaboration, and transforming industries in the coming years.
The next step in the evolution of specialized hardware for AI is rooted in addressing the performance efficiency loss from data movement between computational units and memory. This can be achieved through analog in-memory computing which eliminates the Von Neuman bottleneck and allows highly-parallel computations directly in memory using memristive crossbar arrays. Although Memristive crossbar arrays are a promising future Analog technology for accelerating AI workloads, their inherent noise and non-idealities demand for improved algorithmic solutions.
We introduce the IBM Analog Hardware Acceleration Kit, a first of a kind open source toolkit to simulate crossbar arrays from within PyTorch, to conveniently estimate the impact of material properties and non-idealities on the accuracy for arbitrary ANNs (freely available at https://github.com/IBM/aihwkit). This platform allows understanding, evaluating, and experimenting with emerging analog AI accelerators. Our roadmap and capabilities include algorithmic innovations from IBM Research around hardware-aware training, mixed-precision training, advanced analog training optimizers using parallel rank-update in analog, and allowing inference on real research Phase-change memory (PCM)-based analog AI chip prototypes, as well as allowing the research community to extend the toolkit with new devices, analog presets, algorithms, etc.
We will show an interactive demo of how the toolkit can be used online though our web front-end cloud composer. The composer provides a set of templates and a no-code experience to introduce the concepts of analog AI, configure experiments, and launch training experiments. We are actively working to include inference experiments in simulation and a real PCM-based analog AI chip.