Learning mechanisms of uncertainty and neuromodulation

Learning systems are, by default, adaptive. Experience shapes the parame- ters of artificial systems, as well as it changes the connectivity of biological brains. Nonetheless, our attempts to create artificial learning systems have shown that continuous learning leads to overfitting recent data at t...

Descripción completa

Detalles Bibliográficos
Autor: Puigbò Llobet, Jordi-Ysard
Tipo de recurso: tesis doctoral
Estado:Versión publicada
Fecha de publicación:2019
País:España
Institución:CBUC, CESCA
Repositorio:TDR. Tesis Doctorales en Red
OAI Identifier:oai:www.tdx.cat:10803/667643
Acceso en línea:http://hdl.handle.net/10803/667643
Access Level:acceso abierto
Palabra clave:Learning
Neuromodulation
Uncertainty
Neocortex
Computational Neuroscience
Aprenentatge
Neuromodulació
Incertesa
Neurociència computacional
62
Descripción
Sumario:Learning systems are, by default, adaptive. Experience shapes the parame- ters of artificial systems, as well as it changes the connectivity of biological brains. Nonetheless, our attempts to create artificial learning systems have shown that continuous learning leads to overfitting recent data at the ex- pense of the older. While the field compensated this loss by segregating training from exploitation phases, this comes at the cost of sacrificing the adaptation to uncertain or new situations. How do animals robustly forage for food, find their lairs or flee from predators in ever-changing conditions and, sometimes unfamiliar situations? This dissertation proposes that our brains flexibly change between learning modes, favoring exploitation of previous knowledge or the incorporation/adaptation of new one. From the perspective of fine-tuning perception, this thesis presents a framework to unveil some of the mechanisms that biology can use to learn from uncertain situations rapidly. First, we identify two components of rapid learning by exploring how learning speed can be modulated not just explicitly (i.e., changing a learning rate parameter) but also implicitly (i.e., changing network dynamics) by the modulation of recurrent inhibitory networks. Studying the interactions of cholinergic neuromodulation with local and global inhibition allows us to differentiate between two operation modes that switch between robust exploitation of existing representations and flexibly exploring potential alternatives. To disambiguate the learning mechanisms behind this learning mode switching by a neuromodulator like acetylcholine, we take a step back and propose a neural model to estimate the input uncertainty. The resulting dynamical system minimizes the squared error relative to the input variance, as a proxy of how much an input was unexpected. We show how this kind of system uses two forms of inhibitory populations to estimate the input, and modulate the learning speed, in synthetic datasets and machine learning benchmarks. Altogether, this model illustrates a neural microcircuit, capable of flexibly incorporating new evidence when inputs are unexpected, facilitating learn- ing speed and providing a mechanism to externally regulating learning speed implicitly.