Generation of short-term memory for items and order by unsupervised learning

Abstract

Event Abstract Back to Event Generation of short-term memory for items and order by unsupervised learning In a multiple-item sequential memory task, animals can remember the identity of successive items and their order of appearance [1]. Such ability puts demands on standard models of short-term memory arising from persistent activity of neurons, which typically encode either the first stimulus or the most recent stimulus. In this presentation we show how repeated presentations of stimulus sequences to a network, with an initially limited, transient response to stimuli, can generate the necessary connectivity pattern to solve a two-item sequential memory task. We find that essential plasticity mechanisms within the network are: 1) Hebbian plasticity, which we implement through a well-characterized triplet form of spike-timing dependent plasticity (STDP) [2]. STDP in a sparse network generates an increase in recurrent excitation that is essential to produce persistent activity and memory. 2) Long-term potentiation of inhibition [3] (LTPi), which increases inhibition from active interneurons to inactive pyramidal cells. LTPi produces the cross-inhibition necessary to prevent memory activity from spreading, thus maintaining the specificity of memory. 3) Homeostasis by multiplicative scaling of synaptic strengths [4] prevents a single attractor state from dominating the network, allowing for multiple, approximately equally visited states to be maintained in the network. We produce a simplified, intuitive example of sequential memory as a bump of activity, which requires local excitation to persist following a transient stimulus and which moves toward neurons tuned to subsequent stimuli, provided some longer-range connections exist. Thus we suggest that such a small-world architecture in the pattern of connections between neurons is essential to generate multiple-item sequential memory.