Conferencia del Prof. Dr. Thiago Mosqueiro

[:es]Título: Fast and stable discrimination in divergent-convergent neural networks: from Deep Learning back to Neuroscience
Ponente: Thiago Mosqueiro
Fecha: 28/11/2016 12:30 h
Lugar: Sala de Seminarios, Edificio Torretamarit
Resumen:
La toma de decisiones en el cerebro se modeliza a menudo mediante una red neuronal divergente-convergente en la que la información sensorial atraviesa primero una capa en la que se realiza el reconocimiento de patrones y la codificación del estímulo (capa de conectividades divergentes), seguida de una capa en la cual se filtra la información (capa de conectividades convergentes), antes de llegar a la capa o capas motoras en las que se toma la decisión. Este mecanismo introduce un retardo entre el input sensorial y la respuesta que es inconsistente con la rapidez necesaria en un comportamiento adaptativo. En la charla se hablará de un reciente mecanismo de retroalimentación que es robusto, favorece codificación dispersa (sparse coding) y acelera la transferencia de información a través de las capas. También se discutirán varios ejemplos que ilustran este mecanismo.
Breve Bio:
Thiago Mosqueiro obtuvo el título de doctor en Ciencias Físicas en 2015 por la Universidad de Sao Paulo. Sus intereses científicos se centran en la física estadística, modelos matemáticos, neurociencias computacionales y machine learning. Actualmente compagina su labor investigadora con la docencia en el Rady School of Management de UCSD, y colabora con grupos de investigación de la Universidad de California en Los Angeles y de la Arizona State University.[:en]Title: Fast and stable discrimination in divergent-convergent neural networks: from Deep Learning back to Neuroscience
Speaker: Thiago Mosqueiro
Date: 28/11/2016 12:30 h
Location: Sala de Seminarios, Edificio Torretamarit
Abstract:
The coding basis for decision making is often provided by a minimal number of higher-order neurons. Before reaching premotor decision layers, sensory information travels through several neural layers. This multi-layered organization is often composed of (i) divergent connectivities, which are essential for pattern recognition and stimulus codification, and (ii) convergent connectivities, which filter down information. However, this architecture based on multiple neuronal layers induces a time lag between peripheral input and adaptive behavior (output), which is inconsistent with the need for speed. Furthermore, an accentuated divergent-convergent architecture may also amplify noise and generate unstable dynamics, which impairs the sensory representation of external stimuli. In this talk, we discuss a recent feedback mechanism that presents robust gain-control, sustains sparse coding, and accelerates the information transfer through layers. An example of such synaptic organization is the early olfactory processing stage of all insects, the Mushroom Bodies (MBs), where a strong divergence from 2k to 300k neurons is followed by a convergence to only 400 neurons. The stability analysis of this system provided an analytical formula for the gain-control maintenance. In addition to in-vivo recordings, we used data from gas-sensor arrays to show that this architecture learns more complex spatio-temporal patterns. Moreover, we will use data from gas-sensor arrays to motivate pre-training of such network. Thus, because such connectivities are ubiquitous to many brains, we believe divergent-convergent networks play a central role in stable and fast decision-making processes in the brain.
Brief Bio:
Dr. Mosqueiro graduated in Physics (2008) and got his PhD in Physics (2015) from the University of São Paulo, with an internship (2014) at the BioCircuits Institute and Rady School of Management, both part of the University of California San Diego. His main areas of interest are statistical physics, mathematical modeling, computational neuroscience, and machine learning. Thiago is currently a Post Doc at the BioCircuits Institute, working in collaboration with the University of California Los Angeles and the Arizona State University, and a lecturer at the Rady School of Management.[:]