I would like to propose an approach to the science of deep learning that roughly follows what physicists do to understand reality: (1) explore phenomena through controlled experiments, and (2) build theories based on toy mathematical models and non-fully- rigorous mathematical reasoning. I will illustrate (1) with the LEGO study (LEGO stands for Learning Equality and Group Operations), where we observe how transformers learn to solve simple linear systems of equations. I will also briefly illustrate (2) with an analysis of the emergence of threshold units when training a two-layers neural network to solve a simple sparse coding problem. The latter analysis connects to the recently discovered Edge of Stability phenomenon.

Based on joint works with Kwangjun Ahn, Arturs Backurs, Sinho Chewi Ronen Eldan, Suriya Gunasekar, Yin Tat Lee, Felipe Suarez, Tal Wagner, Yi Zhang, see arxiv.org/abs/2206.04301 and arxiv.org/abs/2212.07469.

Physics of AI — some first steps

Most work on scholarly document processing assumes that the information processed is trustworthy and factually correct. However, this is not always the case. There are two core challenges, which should be addressed: 1) ensuring that scientific publications are credible — e.g. that claims are not made without supporting evidence, and that all relevant supporting evidence is provided; and 2) that scientific findings are not misrepresented, distorted or outright misreported when communicated by journalists or the general public. In this talk, I will present some first steps towards addressing these problems, discussing our research on exaggeration detection, scientific fact checking, and on modelling information change in scientific communication more broadly.

Beyond Fact Checking — Modelling Information Change in Scientific Communication

How do we communicate with AI Agents that learn? One obvious answer is via language. Indeed, humans have evolved languages over tens of thousands of years to provide useful abstractions for understanding and interacting with each other and with the physical world. The claim advanced by some is that language influences how we think, what we perceive, how we focus our attention, and what we remember. We use language to capture and share our understanding of the world around us, to communicate high-level goals, intentions and objectives, and to support coordination with others. Importantly, language can provide us with useful and purposeful abstractions that can help us to generalize and transfer knowledge to new situations.

Language comes in many forms. In Computer Science and in the study of AI, we have historically used formal knowledge representation languages and programming languages to capture our understanding of the world and to communicate unambiguously with computers. In this talk I will discuss how formal language can help agents learn and reason with a deep dive on one particular topic – reinforcement learning. I’ll show how we can exploit the syntax and semantics of formal language and automata to aid in the specification of complex reward-worthy behavior, to improve the sample efficiency of learning, and to help agents learn what to remember. In doing so, formal language can help us address some of the challenges to reinforcement learning in the real world.

(Formal) Languages Help AI Agents Learn and Reason

Reinforcement Learning provides an attractive suite of online learning methods for personalizing interventions in Digital Behavioral Health. However, after a reinforcement learning algorithm has been run in a clinical study, how do we assess whether personalization occurred? We might find users for whom it appears that the algorithm has indeed learned in which contexts the user is more responsive to a particular intervention. But could this have happened completely by chance? We discuss some first approaches to addressing these questions

We used Reinforcement Learning; but did it work?

Anima will present exciting developments in the use of AI for scientific applications. This includes diverse domains such as weather and climate modeling, deep earth modeling, genome modeling, etc. We have developed principled approaches that enables zero-shot generalization beyond the training domain. This includes neural operators that yield 4-5 orders of magnitude speedups over numerical weather models and other scientific simulations. They learn mappings between function spaces that makes them ideal for capturing multi-scale processes.

AI accelerating Science: Neural Operators for Learning on Function Spaces

Consider an intelligent virtual assistant such as Siri, or perhaps a more capable future version of it. Should we think of all of Siri as one big agent? Or is there a separate agent on every phone, each with its own objectives and/or beliefs? And what should those objectives and beliefs be? Such questions reveal that the traditional, somewhat anthropomorphic model of an agent – with clear boundaries, centralized belief formation and decision making, and a clear given objective – falls short for thinking about today’s AI systems. We need better methods for specifying the objectives that these agents should pursue in the real world, especially when their actions have ethical implications. I will discuss some methods that we have been developing for this purpose, drawing on techniques from preference elicitation and computational social choice. But we need to specify more than objectives. When agents are distributed, systematically forget what they knew before (say, for privacy reasons), can be simulated by others, and potentially face copies of themselves, it is no longer obvious what the correct way is even to do probabilistic reasoning, let alone to make optimal decisions. I will explain why this is so and discuss our work on doing these things well. (No previous background required.)

New Design Decisions for Modern AI Agents

The development of robots that can learn to interact with the world and manipulate the objects in it has emerged as one of the greatest and so far largely unsolved challenges in robotics research. In this talk, I will argue that the development of such advanced machines requires a transition from classical manual design with purely model-based control to a novel synthesis paradigm. We need to allow the machine to autonomously develop its own blueprint and algorithmically generate its topological, kinematic, and dynamic self. Building on this, it shall develop controls for its own body as it moves, learns to manipulate objects in a controlled way, and sensitively interacts with the world.

Drawing from our work in torque-controlled lightweight robots towards human-safe tactile robots that can manipulate, fly, or drive, I outline the technological quantum leaps that recently have taken place. In particular, this progress was made possible by a human-centered design, soft and force-sensitive control, contact reflexes, and model-based machine learning. In the real world, by enabling human-robot coexistence, collaboration, and interaction for the first time, this robotic technology has proven transformative to traditional manufacturing already around the globe. Increasingly, it is now impacting professional services, domestic applications, medicine and healthcare.

After that, I will use our current work to chart the path toward the next generation of tactile machines. We have taken first steps towards autonomously designing and building machines that have the ability to learn their self and thus adapt to changes in body topology and ultimately their entire dynamics. Finally, I will present recent results on designing modular control and learning architectures that achieve complex behaviors for challenging manipulation problems while being provably stable.

Robots with a sense of touch: self-replicating the machine and learning the self

Computer vision is one of the great AI success stories. Yet we are still far from having machine systems that can reliably and robustly see everything a human being sees in an image or in the real world. Despite rapid advances in self-supervised visual and multimodal representation learning, we are also far from having systems that can learn to see as richly as a human does, from so little data, or that can learn new visual concepts or adapt their representations as quickly as a human does. And even today’s remarkable generative image synthesis systems imagine the world in a very different and fundamentally less flexible way than human beings do. How can we close these gaps? I will describe several core insights from the study of human vision and visual cognitive development that run counter to the dominant trends in today's computer vision and machine learning world, but that can motivate and guide an alternative approach to building practical machine vision systems.

Technically, this approach rests on advances in differentiable and probabilistic programming: hybrids of neural, symbolic and probabilistic modeling and inference that can be more robust, more flexible and more data-efficient than purely neural approaches to learning to see. New probabilistic programming platforms offer to make these approaches scalable as well. Conceptually, this approach draws on classic proposals for understanding vision as ‘inverse graphics’, ‘analysis by synthesis’ or ‘inference to the best explanation’, and the notion that at least some high-level architecture for scene representation is built into the brain by evolution rather than learned from experience, reflecting invariant properties of the physical world. Learning then enables, enriches and extends these built-in representations; it does not create them from scratch. I will show a few examples of recent machine vision successes based on these ideas, from our group and others. But the hardest problems are still very open. I will highlight some ‘Grand Challenge’ tasks for building machines that learn to see like people: problems that far outstrip the abilities of any current system, and that I hope can inspire the next steps towards progress for computer vision researchers regardless of which approach they favor.

Learning to See the Human Way

Robert S. Engelmore Memorial Award Lecture

I share insights on the interaction of humans and AI to jointly solve end-to-end complex problems. The talk is based on my experience with research and engineering of real autonomous mobile service robots, and AI in the finance domain. I will illustrate the integration of perception (data), cognition (reasoning and learning), and action (execution and feedback). And I will focus on several components, including AI to effectively discover and standardize data, for behavior understanding, to simulate and generate synthetic data, to learn from data, principles, and experience, and to provide trustful explanations.

Symbiotic Human-AI Interaction: Experience-Based Insights from AI in Robotics and AI in Finance

AAAI Award for Artificial Intelligence for the Benefit of Humanity

I will share experiences from working on organ exchanges for the last 18 years, ranging from market designs to new optimization algorithms to large-scale fielding of the techniques and even to computational policy optimization. Originally in kidney exchange, patients with kidney disease obtained compatible donors by swapping their own willing but incompatible donors. I will discuss many modern generalizations of this basic idea. For one, I will discuss never-ending altruist donor chains that have become the main modality of kidney exchanges worldwide and have led to over 10,000 life-saving transplants. Since 2010, our algorithms have been running the national kidney exchange for United Network for Organ Sharing, which has grown to include 80% of the transplant centers in the US. Our algorithms autonomously make the transplant plan each week for that exchange, and have been used by two private exchanges before that. I will summarize the state of the art in algorithms for the batch problem, approaches for the dynamic problem where pairs and altruists arrive and depart, techniques that find the highest-expected-quality solution under the real challenge of unforeseen pre-transplant incompatibilities, algorithms for pre-match compatibility testing, and approaches for striking fairness-efficiency tradeoffs. I will describe the FUTUREMATCH framework that combines these elements and uses data and supercomputing to optimize the policy from high-level human value judgments. The approaches therein may be able to serve as ways of designing policies for many kinds of complex real-world AI systems. I will also discuss the idea of liver lobe exchanges and cross-organ exchanges, and how they have started to emerge for real.

Modern Organ Exchanges: Market Designs, Algorithms, and Opportunities

Recent development in the field of explainable artificial intelligence (XAI) has helped improve trust in Machine-Learning-as-a-Service (MLaaS) systems, in which an explanation is provided together with the model prediction in response to each query. However, XAI also opens a door for adversaries to gain insights into the black-box models in MLaaS, thereby making the models more vulnerable to several attacks. For example, feature-based explanations (e.g., SHAP) could expose the top important features that a black-box model focuses on. Such disclosure has been exploited to craft effective backdoor triggers against malware classifiers. To address this trade-off, we introduce a new concept of achieving local differential privacy (LDP) in the explanations, and from that we establish a defense, called \textsc{XRand}, against such attacks. We show that our mechanism restricts the information that the adversary can learn about the top important features, while maintaining the faithfulness of the explanations.

XRand: Differentially Private Defense against Explanation-Guided Attacks

Probabilistic models based on continuous latent spaces, such as variational autoencoders, can be understood as uncountable mixture models where components depend continuously on the latent code. They have proven expressive tools for generative and probabilistic modelling, but are at odds with tractable probabilistic inference, that is, computing marginals and conditionals of the represented probability distribution. Meanwhile, tractable probabilistic models such as probabilistic circuits (PCs) can be understood as hierarchical discrete mixture models, which allows them to perform exact inference, but often show subpar performance in comparison to continuous latent-space models. In this paper, we investigate a hybrid approach, namely continuous mixtures of tractable models with a small latent dimension. While these models are analytically intractable, they are well amenable to numerical integration schemes based on a finite set of integration points. With a large enough number of integration points the approximation becomes de-facto exact. Moreover, using a finite set of integration points, the approximation method can be compiled into a PC performing 'exact inference in an approximate model'. In experiments, we show that this simple scheme proves remarkably effective, as PCs learned this way set new state-of-the-art for tractable models on many standard density estimation benchmarks.

Underline digital video library

Physics of AI — some first steps

Beyond Fact Checking — Modelling Information Change in Scientific Communication

(Formal) Languages Help AI Agents Learn and Reason

We used Reinforcement Learning; but did it work?

AI accelerating Science: Neural Operators for Learning on Function Spaces

New Design Decisions for Modern AI Agents

Robots with a sense of touch: self-replicating the machine and learning the self

Learning to See the Human Way

Symbiotic Human-AI Interaction: Experience-Based Insights from AI in Robotics and AI in Finance

Modern Organ Exchanges: Market Designs, Algorithms, and Opportunities

XRand: Differentially Private Defense against Explanation-Guided Attacks

Continuous Mixtures of Tractable Probabilistic Models

Stay up to date with the latest Underline news!

PRESENTATIONS

CONFERENCES

COMPANY

RESOURCES