Talks are sometimes moved if there is a seminar

https://math.ethz.ch/sfs/news-and-events/young-data-science.html?s=fs21

Title: Neural Networks and Multiscale Entropies (Zoom https://ethz.zoom.us/j/511336187)
Abstract: One of the most striking properties of our world is the enormous variety of length scales in its structure. Complex physical phenomena and real-world signals and images contain structures at different scales. In this talk, we study distributions that, given a mean constraint, simultaneously maximize entropies at different scales. This is motivated by the goal of analyzing neural networks while exploiting their distinguishing characteristic, that is, their multi-level architecture. Neural networks have found remarkable success in many areas of science such as image classification, yet finding a proper theoretical ground for designing their architecture and training procedure remains an active research area. In this talk, we define the notion of multiscale entropic regularization and demonstrate how it naturally applies to the multilevel architecture of neural networks. We then propose simulating a multiscale extension of the Gibbs-Boltzmann distribution as a training procedure that is an alternative to stochastic gradient descent and is based on the chain rule of entropy rather than the chain rule of derivatives. This information-theoretic approach is inspired by both multiscale techniques of chaining in high-dimensional probability and renormalization group theory in statistical physics. It can also be viewed as an extension of the maximum entropy results of Jaynes (1957) and Csiszár (1975) and a Bayesian expansion of the renormalization group.

Joint work with Emmanuel Abbe (EPFL).

Bio:
Amir R. Asadi is a final year Ph.D. candidate in the Department of Electrical and Computer Engineering at Princeton University where he is advised by Emmanuel Abbe. He received B.Sc. degrees in Electrical Engineering and in Mathematics from Sharif University of Technology in Iran, both in 2015, and M.A. degree in Electrical Engineering from Princeton University in 2017. The recipient of the Anthony Ephremides Fellowship from Princeton University in 2016, his research is on the interface of machine learning, information theory, and high-dimensional probability.

Group Meeting and chat about Free Probability Reading Group

Series of papers on the optimization of spin glasses : https://arxiv.org/abs/1812.04588, https://arxiv.org/abs/2001.00904, https://arxiv.org/abs/2009.11481

DST Offset // FOCM Seminar this day: http://focm-society.org/events.php

https://eth-its.ethz.ch/activities/its-science-colloquium/fan-yang.html

Journal Club: Maria Refinetti (https://arxiv.org/pdf/2102.11742.pdf)

Journal Club: Maria Refinetti (https://arxiv.org/pdf/2102.11742.pdf)

Journal Club: Tsigler (https://arxiv.org/pdf/1906.11300.pdf)

Journal Club: Tsigler (https://arxiv.org/pdf/1906.11300.pdf)

Joel Tropp is giving a colloquium in CS: https://inf.ethz.ch/news-and-events/events/event-detail.joel-tropp-finding-structure-with-randomness.61221.html

Community detection using t-SNE /YDS Seminar (16:30)