Continuous-Time Heteroassociative Memory at Biological Timescales

Marc Gong Bacvanski, Liu Ziyin, Tomaso Poggio

Keywords: Heteroassociative memory, dynamical systems, continuous time, biologically plausible learning

TL;DR: We show that in continuous-time heteroassociative networks where inference and learning coevolve, successful learning is governed by temporal overlap between cue and error signals, and requires plasticity timescales that match biological regimes.

Abstract:

Associative learning in biological systems unfolds continuously in time, yet most models implicitly assume synchronized, discrete updates that instantaneously deliver the correct teaching signal to each synapse. We study a model of continuous-time heteroassociative memory neural networks in which inference and learning coevolve under coupled ODEs, and where different error-propagation topologies determine how error signals reach synapses. Longer error pathways result in propagation delays of error signals, while deeper networks result in propagation delays of cue signals. We experimentally investigate the boundary at which learning begins to fail under different combinations of propagation time constants, synaptic plasticity time constants, error signal delay $\Delta$, and cue duration $T$.
We find that the plasticity time constant must substantially exceed the cue duration ($t_\text{plas}/T\approx10\text{–}50$ in our settings), a regime aligned with biological evidence. Together, these results yield testable predictions for associative memories in neuroscience, and practical design guidance for hardware implementations.

Submission Number: 26