Please join Hector Manuel Lopez Rios, Postdoctoral Fellow, for an upcoming seminar.
Gated Recurrent Neural Networks as a Framework to Study Neuromodulation: improving memory capacity and generalization of dynamical tasks
Recurrent neural networks (RNNs) have become a standard framework for modeling how neural circuits learn and process information, yet most models assume that computation is mediated entirely through synaptic connectivity (short-range communication). However, neurons also communicate extra-synaptically via chemical signaling (e.g., neuropeptides, monoamines). These extra-synaptic interactions are essential for transitions between behavioral states, but their computational role has received limited analytical treatment within the associative memory and RNN frameworks. In this talk, I will introduce a minimal model that treats neuromodulation as dynamical tuning of neuronal integration timescales. We augment a time-continuous Hopfield network with a neuromodulatory gating layer inspired by the neuropeptidergic connectome of C. elegans. Using numerical simulations and dynamical mean-field theory, we show that neuromodulation increases the memory capacity of the network beyond the classical limit and enables generalization to unseen conditions in timing tasks (measure-wait-go). Our results suggest that neuromodulatory signaling is an active computational resource that expands the dynamical repertoire of neural circuits.
Hector Manuel Lopez Rios is a postdoctoral fellow at the University of Chicago working with Suri Vaikuntanathan on the dynamics and information processing of biological and chemical networks. He earned his Ph.D. at Northwestern University, advised by Monica Olvera de la Cruz. Hector holds a bachelor's degree in chemical engineering from UNAM in Mexico City, Mexico.