SINC(I)   25518
INSTITUTO DE INVESTIGACION EN SEÑALES, SISTEMAS E INTELIGENCIA COMPUTACIONAL
Unidad Ejecutora - UE
congresos y reuniones científicas
Título:
Cortical-like dynamics in recurrent circuits optimized for sampling-based probabilistic inference
Autor/es:
AITCHISON, LAURENCE; ECHEVESTE, RODRIGO; LENGYEL, MÁTÉ; HENNEQUIN, GUILLAUME
Lugar:
(online por la pandemia)
Reunión:
Conferencia; Bernstein Conference 2020; 2020
Institución organizadora:
Bernstein Network
Resumen:
The dynamics of sensory cortices show a suite of basic, ubiquitous features that have so farescaped a common, principled theoretical account. These include strong, inhibition-dominatedtransients at stimulus onset, gamma oscillations, and noise variability ? all stimulus-dependent. Wepresent a unifying model in which all these dynamical phenomena emerge as a consequence of theefficient implementation of the same computational function: fast probabilistic inference [1]. For this, we used a novel approach and trained a recurrent E/I neural circuit model of a V1 hypercolumn. The network was required to modulate not only the mean (as conventional) but also the variability of its stationary response distributions in order to match the corresponding input-dependent posterior distributions inferred by an ideal observer. The optimized neural circuit featured a number of remarkable computational and biological properties. Computationally, after training on a reducedstimulus set, it exhibited strong forms of generalization by producing near-optimal responsedistributions to novel inputs which required qualitatively different responses. Furthermore, the network discovered non-equilibrium dynamics, a state-of-the-art machine learning strategy to speed up inferences. The circuit also exhibited realistic biological properties for which it was not trained directly. It achieved divisive normalization and displayed marked transients at stimulus onset, as well as strong gamma oscillations, both scaling with stimulus contrast. Crucially, these dynamical phenomena did not emerge in other control networks trained to match either mean responses only, or mean responses and variances (without modulating response co-variances). Further analyses of transients and oscillations in the optimized network revealed distinct functional roles for them in speeding up inferences and madepredictions that we confirmed in novel analyses of awake monkey V1 recordings. Our results offer a principled theoretical account of the basic motifs of cortical dynamics and predict further properties of these motifs that can be tested in future experiments.