Learning and Retrieval of Sequential Activity Using Temporally Asymmetric Hebbian Learning Rules

Abstract

Sequential activity in brain networks has been observed across multiple species, neural structures, and behavioral contexts. This activity consists of groups of neurons that are active at different times in a reproducible manner, and in many contexts appears to be internally generated, proceeding autonomously in the absence of ongoing sensory input. Such activity underlies a number of motor behaviors and working memory tasks. This activity displays diverse temporal characteristics, with properties that depend on the neural area, and on the timescale of observation, and experimental evidence suggests that it forms gradually over the course of task learning and behavioral training. Understanding how single neuron and network mechanisms give rise to such activity remains a fundamental challenge in neuroscience.Learning in the brain is believed to occur through activity-dependent modifications of synaptic connectivity. We explore a class of synaptic plasticity rules, temporally asymmetric Hebbian learning rules, and investigate whether their application in network models can account for the diversity of temporal characteristics. These Hebbian rules transform a sequence of input patterns into synaptic weight updates. After learning, recalled sequential activity is reflected in the transient correlation of network activity with each of the stored input patterns. We use mean field-theory to develop a low-dimensional description of the network dynamics and compute the storage capacity of these networks. On short timescales, we find that the temporal statistics of this activity are consistent with experimental observations in multiple brain structures. Nonlinearities in the learning rule alter the degree of temporal sparsity during retrieval. On long timescales, we find that modification of synaptic connectivity due to noise, storage of other patterns, or rehearsal of learned patterns, produces sequential activity with highly labile dynamics. In many experimental settings, the speed of sequential activity can be flexibly controlled depending on task context. We find that adding heterogeneity in the learning rule across synapses allows retrieval speed to be controlled through changes in the external input drive. Speed changes are accompanied by a temporal rescaling of activity dynamics, similar to neural dynamics observed in cortex.