Embedded Files
(Times are according the Pacific Time zone)
6.00 AM Welcome & Best Paper Award
6.15 AM Spotlight presentations
6.45 AM Poster Session (Gather.town)
8.15 AM (Invited talk) Kristian Kersting "Neural Probabilistic Circuits are All You Need"
9.00 AM (Invited talk) George Papamakarios "Normalizing Flows"
9.45 AM (Invited talk) Martin Trapp "Bayes Meets Probabilistic Circuits"
10.30 AM Virtual coffee break
11.00 AM (Invited talk) Arthur Choi "Using Tractable Circuits to Explain Statistical Machine Learning"
11.45 AM Spotlight presentations
12.15 PM Poster Session (Gather.town)
1.45 PM (Invited talk) Zelda Mariet "Modeling Negative Dependence at Scale"
2.30 PM (Invited talk) Anima Anandkumar "Tractable Uncertainty Estimation in Deep Learning"
3.15 PM Panel and closing remarks
Invited Speakers
Caltech, NVIDIA
Tractable Uncertainty Estimation in Deep Learning
Current deep learning models tend to be overconfident even when they make errors. Thus, they cannot yet be used in real-world deployments of safety-critical applications like autonomous driving, robotics, and healthcare. I will present tractable and lightweight uncertainty estimation methods that provide calibrated uncertainties even under challenging conditions of distributional shifts such as sim-to-real domain adaptation.
University of California, Los Angeles
Using Tractable Circuits to Explain Statistical Machine Learning
Recent and rapid advances in Artificial Intelligence (AI), particularly in the form of deep neural networks, has opened many new possibilities, but it has also brought with it many new challenges. In particular, it has become increasingly apparent that while deep neural networks are highly performant, they can also be opaque and brittle. We do not have enough understanding of why and when they work well, and why they may fail completely when faced with new situations not seen in the training data. In this talk, we propose a symbolic approach to explaining the behavior and verifying the properties of machine learning models, which is based on sustained advances in logical and probabilistic reasoning. We show how our approach facilitates the analysis of a neural network, helping us to understand its behavior, and in turn, providing directions towards learning better and more robust models
George Papamakarios
George Papamakarios
DeepMind, London
Normalizing flows are an approach to probabilistic modelling that enables tractable probability-density evaluation and sampling. But are they tractable? What applications do they support? What are their strengths and computational limitations? In this talk I will give an overview of normalizing flows and discuss their applications, strengths, weaknesses, and some recent progress.
TU Darmstadt
Neural Probabilistic Circuits are All You Need
Probabilistic circuits (PCs) are a promising avenue for probabilistic modelling, as they permit a wide range of exact but tractable inference routines. Unfortunately, often they still lack the expressive power of intractable models based on deep neural networks. To combine the best of both worlds, I shall introduce and Illustrate neural PCs that allow one to harness the expressive power of neural networks while still maintaining (some) tractability guarantees. Via gating nodes one can establish universal function approximators and causal inference. Via compiling a large number of arithmetic operations in a single monolithic einsum-operation, one can achieve speedups and memory savings of up to two orders of magnitude. Via the fourier transformation, one can realize deep liklihoods for time series. This all illustrates the power of hybrid models.
Aalto University
Bayes Meets Probabilistic Circuits
The Bayesian approach provides a powerful way to handle and manipulate uncertainties and is used in many scientific disciplines to solve real-world problems. Consequently, recent work on tractable probabilistic models has shown several promising ways of using the Bayesian approach, e.g., for Bayesian structure learning. Alternatively, tractable probabilistic models can enable more efficient Bayesian inference, e.g., by using an expert-based approximation of Gaussian processes. In this talk, I will give a systematic overview of Bayesian approaches in the literature on probabilistic circuits, discuss some recent work in the field, and point out potential research avenues that the two fields offer for each other.
MIT, Google AI
Modeling Negative Dependence at Scale
Negative dependence provides a rich mathematical framework through which machine learning problems that require a theoretical model of diversity can be analyzed; such problems include experimental design, sensor placement, active learning, and recommender system tasks. This talk will present scalable sampling and learning algorithms for popular negatively dependent measures such as determinantal point processes (DPPs) and volume sampling. We will discuss how definitions of negative dependence can be relaxed via exponentiated Strongly Rayleigh (eSR) measures and completely log-concave (CLC) distributions, and how these generalizations can maintain scalability guarantees. Finally, we will present some remaining open questions regarding tractable representations for models of diversity.
Google Sites
Report abuse