Abstract

In this talk I will first give a brief broad overview of my current research lines after leaving CBL , to then focus on the use of machine learning tools in order to model cortical visual processing in terms of Bayesian inference. I will first recap previous work at CBL where we showed how recurrent neural networks trained for fast sampling-based Bayesian inference display stereotypical features of cortical dynamics, such as transients and oscillations. This work relied on counting with a generative model for the domain of interest, which was then inverted to obtain an ideal observer model, which the network was asked to mimic. In current work we learn both the generative and inference models from the data by use of variational auto-encoders. I’ll show ongoing work on this line, particularly on how to extend classical VAEs to obtain models of visual perception with well behaved uncertainty estimates. I will also discuss how these extensions can prove useful in other classical machine learning domains.

Related papers:

Echeveste et al. (2020), Cortical-like dynamics in recurrent circuits optimized for sampling-based probabilistic inference. https://www.nature.com/articles/s41593-020-0671-1

Catoni et al. (2024), Uncertainty in latent representations of variational autoencoders optimized for visual tasks. https://arxiv.org/abs/2404.15390

Short bio:

Rodrigo Echeveste obtained his Bachelor and Masters Degrees in Physics from Balseiro Institute in Argentina, and his PhD from the Goethe University of Frankfurt, Germany. He then did a three-year postdoc at CBL . Rodrigo currently holds a permanent research position as Adjunct Researcher from Argentina’s National Research Council (CONICET) at the Research Institute for Signals, Systems and Computational Intelligence, sinc(i), and is an Adjunct Professor at the National University of Litoral (UNL). His work lies at the intersection of Computational Neuroscience and Machine Learning. Rodrigo currently serves as Secretary of Argentina’s Society for Neuroscience Research (SAN).