Seminario impartido por Miguel Ibañez-Berganza (IMT Lucca, Italia).
Sensorial neural codes often present information-limiting noise correlations (ILNC), or across-neuron correlations that limit the amount of information (of a stimulus) that a neural populations can encode. Such information-limiting correlations have long been though to be detrimental for behavioural performance, but recent experiments have shown that they can actually enhance it [1,2]. To explain such fact, as well as the evolutionary origin of ILNC, Panzeri and collaborators have proposed a general theoretical mechanism called coincidence-detection [2]. According to coincidence-detection, if x(s) is the activity of a neural population encoding a stimulus s, correlations between the components of x may enhance the transmitted amount of stimulus information to downstream, readout neural populations y(x). ILNC in x may, in other words, enhance the stimulus mutual information I(s,y) despite reducing I(s,x).
I will present an approximate solution of a model of coincidence-detection in a simple three-layer (s->x->y) biologically-inspired neural network. Such solution stands on the firing properties of a Leaky Integrate-and-Fire (LIF) model of spiking neuron with colored-noise input current. The analysis of this model represents an instance of the challenging problem of computing first passage times of non-Markovian stochastic processes. The solution eventually provides us with an interpretable, intuitive explanation of the apparently paradoxical effect of coincidence-detection.
[1] Panzeri et al. Nat Rev Neurosci 23, 551–567 (2022)
[2] Valente et al. Nat Neurosci 24, 975–986 (2021)
- Fecha: Jueves 5 de Octubre de 2023
- Lugar: Facultad de Ciencias – Laboratorio de Física Computacional, Departamento de Electromagnetismo y Física de la Materia, Planta Baja
- Horario: 11.30 horas
- Organiza: Departamento de Electromagnetismo y Física de la Materia
- Más información: serena@onsager.ugr.es