Turbo: A Physical-Minded Approach to Generalized Autoencoders

<!--HTML--><p>This presentation explores the interconnection between the information bottleneck and a new framework called Turbo.<br>We will first formulate a variational approximation of the information bottleneck and show how several existing models can be seen as particular cases. We then address the limitations of the information bottleneck in physical problems and propose the Turbo framework as a solution.<br>Turbo is a generalized autoencoder framework that maximizes the mutual information between the input and output of the encoder and decoder.<br>The framework allows for the interpretation and creation of diverse models, as well as the choice of encoder and decoder architecture.<br>The application of Turbo to several problems will be demonstrated, including collider physics generation, image-to-image translation, and inverse problems in astronomy.</p><p><i>Slava Voloshynovskiy &nbsp;received a radio engineer degree from Lviv Polytechnic Institute, Lviv, Ukraine, in 1993 and a Ph.D. degree in electrical engineering from the State University Lvivska Polytechnika, Lviv, Ukraine. Since 1999, he has been with the University of Geneva, Switzerland, where he is currently a Professor with the Department of Computer Science and head of the Stochastic Information Processing group. His research interests are in information-theoretic aspects of stochastic image modeling, digital watermarking, physical uncloneable functions and machine learning that includes generative models, digital twins and anomaly detection.&nbsp;</i></p><p><strong>Coffee will be served at 10:30.</strong></p>