Starts 13 Feb 2026 11:00
Ends 13 Feb 2026 12:00
Central European Time
QLS Seminar - Memory Capacity, Novelty Encoding, and Continual Learning in Biologically Plausible Neural Networks
ICTP
Common Area, Old SISSA building Second floor
Via Beirut, 2
I will present two projects and briefly discuss my interest in continual learning.
The first project is a theoretical study of memory capacity in heterogeneous neural networks. We generalize classical Hopfield-model results to networks with heterogeneous neuron coding levels and arbitrary connectivity, deriving an analytical expression for maximal memory capacity. Although heterogeneity typically reduces capacity, we show that capacity can be preserved when coding levels and in-degree are positively correlated. This principle holds across multiple biologically relevant settings and can be experimentally tested. In addition, in a two-layer network representing the hippocampal CA3–DG circuit, we find that a quasi-indexing encoding scheme in the DG maximizes capacity, extending the hippocampal index theory of memory.
The second project is an experimental study of how the basal forebrain encodes salience and novelty. We identify a population of basal forebrain neurons that exhibit phasic bursting responses to surprising external events, including unexpected rewards, novel stimuli, and sequence violations. These neurons do not encode reward prediction errors, but instead signal surprise and salience, suggesting a mechanism by which the basal forebrain primes cortical circuits for enhanced processing of important information.
The first project is a theoretical study of memory capacity in heterogeneous neural networks. We generalize classical Hopfield-model results to networks with heterogeneous neuron coding levels and arbitrary connectivity, deriving an analytical expression for maximal memory capacity. Although heterogeneity typically reduces capacity, we show that capacity can be preserved when coding levels and in-degree are positively correlated. This principle holds across multiple biologically relevant settings and can be experimentally tested. In addition, in a two-layer network representing the hippocampal CA3–DG circuit, we find that a quasi-indexing encoding scheme in the DG maximizes capacity, extending the hippocampal index theory of memory.
The second project is an experimental study of how the basal forebrain encodes salience and novelty. We identify a population of basal forebrain neurons that exhibit phasic bursting responses to surprising external events, including unexpected rewards, novel stimuli, and sequence violations. These neurons do not encode reward prediction errors, but instead signal surprise and salience, suggesting a mechanism by which the basal forebrain primes cortical circuits for enhanced processing of important information.