## Select Seminar Series

All seminarsYou are here

- Home
- ›
- Studying at the Faculty
- ›
- Seminars ›
- Upcoming Seminars

# Upcoming Seminars

ThursdayDec 15, 202212:15

Vision and AIRoom 1

Speaker:Yossi GandelsmanTitle:Test-time Training with Self-SupervisionAbstract:opens in new windowin html pdfopens in new windowTest-Time Training is a general approach for improving the performance of predictive models when training and test data come from different distributions. It adapts to a new test distribution on the fly by optimizing a model for each test input using self-supervision before making the prediction. This method improves generalization on many real-world visual benchmarks for distribution shifts.
In this talk, I will present the recent progress in the test-time training paradigm. I will show how masked auto-encoding overcomes the shortcomings of previously used self-supervised tasks and improves results by a large margin. In addition, I will demonstrate how test-time training extends to videos - instead of just testing each frame in temporal order, the model is first fine-tuned on the recent past before making a prediction and only then proceeding to the next frame.

ThursdayDec 15, 202213:30

Geometric Functional Analysis and Probability SeminarRoom 155

Speaker:Yeor HafoutaTitle:A Berry-Esseen theorem in L^p under weak dependenceAbstract:opens in new windowin html pdfopens in new windowThe classical Berry-Esseen theorem provides optimal rates in the central limit theorem (CLT) for partial sums of iid random variables. Since then there have been many extensions to ``weakly dependent" (aka mixing) random variables. A related question is the accuracy of approximation by Guassians in L^p (after coupling). The validity of such optimal L^p rates was an open problem by E. Rio, which was recently (2018) solved by S. Bobkov for independent random variables.
In the talk I will present recent results concerning optimal CLT rates in L^p for a variety of weakly dependent random variables like (inhomogeneous) Markov chains, products of random matrices and partially hyperbolic dynamical systems. We will also discuss relations with almost sure rates of approximation by Gaussian random variables (i.e. rates in the almost sure invariance principle).
Edgeworth expansions provide better than optimal CLT rates, with appropriate correction terms.
Our proofs require non-uniform versions of such expansions for weakly dependent random variables, which are of independent interest and have several other applications. In particular, we obtain a non-uniform Berry-Esseen theorem for weakly dependent random variables.

SundayDec 18, 202211:00

Faculty SeminarRoom 155

Speaker:Brit Youngmann Title:Data Tools for Accelerated Scientific DiscoveriesAbstract:opens in new windowin html pdfopens in new windowLike general data scientists, scientists conducting empirical research rely on data and analyze it to extract meaningful insights. Yet, scientists have two qualities that distinguish them from general data scientists: (1) they rely on extensive scientific domain knowledge to make scientific discoveries, and (2) they aim to explain and understand, not simply predict, the real world. These two qualities must be reflected in their data analysis tools to assist them and accelerate the process of making real-world discoveries. In this talk, I will present data tools for accelerated scientific discoveries. In particular, I will present tools that assist scientists in investigating their data using scientific knowledge and helping scientists acquire missing data and domain-specific knowledge required to understand real-world mechanisms and draw trustworthy conclusions.

MondayDec 19, 202211:15

Foundations of Computer Science SeminarRoom 155

Speaker:Or Zamir Title:Algorithmic Applications of Hypergraph and Partition ContainersAbstract:opens in new windowin html pdfopens in new windowWe present a general method to convert algorithms into faster algorithms for almost-regular input instances. Informally, an almost-regular input is an input in which the maximum degree is larger than the average degree by at most a constant factor. This family of inputs vastly generalizes several families of inputs for which we commonly have improved algorithms, including bounded-degree inputs and random inputs. It also generalizes families of inputs for which we don't usually have faster algorithms, including regular-inputs of arbitrarily high degree and all very dense inputs. We apply our method to achieve breakthroughs in exact algorithms for several central NP-Complete problems including k-SAT, Graph Coloring, and Maximum Independent Set.
Our main tool is the first algorithmic application of the relatively new Hypergraph Container Method (Saxton and Thomason 2015, Balogh, Morris and Samotij 2015). This recent breakthrough, which generalizes an earlier version for graphs (Kleitman and Winston 1982, Sapozhenko 2001), has been used extensively in recent years in extremal combinatorics. An important component of our work is a new generalization of (hyper-)graph containers to Partition Containers.

TuesdayDec 20, 202211:00

Special Guest SeminarRoom 1

Speaker:Dimitri KanevskyTitle:Non-associative Moufang loops of point classes on cubic surfacesAbstract:opens in new windowin html pdfopens in new windowIn this talk, I will construct an example of a non-associative Moufang loop of point classes on a cubic surface over a local field and describe a class of cubic surfaces over number fields for which I conjecture that Moufang loops associated with them are non-associative. The question about the existence of non-associative loops of point classes on cubic surfaces was stated in Yuri I. Manin's book "Cubic Forms" about fifty years ago. All required concepts will be recalled.

WednesdayDec 21, 202211:15

Machine Learning and Statistics SeminarRoom 1

Speaker:Yonathan EfroniTitle:Reinforcement Learning in the presence of irrelevant informationAbstract:opens in new windowin html pdfopens in new windowReinforcement Learning (RL) is a field concerned with designing general purpose learning algorithms that solve sequential-decision tasks. In recent years, by using deep neural networks, RL algorithms were applied on high-dimensional and challenging domains, witnessing unprecedented success. Yet, despite recent advancements, the theoretical foundations of high-dimensional RL are not fully understood.
A recurring theme in high-dimensional RL is the presence of irrelevant information in the observations. E.g., in a visual navigation task the observation might capture the movement of clouds, which is irrelevant for reaching the goal location. This calls for natural questions: Can such tasks be learned efficiently, depending only on the complexity of the relevant information? Can RL algorithms be robust to noise in observations? Surprisingly, contemporary RL algorithms may provably fail in the presence of irrelevant information.
In this talk, I will elaborate on these failure cases and present our new provable approaches for high-dimensional RL with irrelevant information. Shared to these are techniques to filter the irrelevant information while guaranteeing near-optimal behavior. I will conclude with experimental results showcasing challenges and solutions in practice.
Bio:
Yonathan is a research scientist at Meta. Prior to that he completed his post-doctorate in Microsoft Research, New York. He obtained his PhD from the Technion, advised by Prof. Shie Mannor, and his Master from the Weizmann institute in Physics. His work won the outstanding paper award in AAAI19 and a best paper award in the OptRL workshop in NeurIPS19.