A Spiking Working Memory Model Based on Hebbian Short-Term Potentiation

Abstract

A dominant theory of working memory (WM), referred to as the persistent activity hypothesis, holds that recurrently connected neural networks, presumably located in the prefrontal cortex, encode and maintain WM memory items through sustained elevated activity. Reexamination of experimental data has shown that prefrontal cortex activity in single units during delay periods is much more variable than predicted by such a theory and associated computational models. Alternative models of WM maintenance based on synaptic plasticity, such as short-term nonassociative (non-Hebbian) synaptic facilitation, have been suggested but cannot account for encoding of novel associations. Here we test the hypothesis that a recently identified fast-expressing form of Hebbian synaptic plasticity (associative short-term potentiation) is a possible mechanism for WM encoding and maintenance. Our simulations using a spiking neural network model of cortex reproduce a range of cognitive memory effects in the classical multi-item WM task of encoding and immediate free recall of word lists. Memory reactivation in the model occurs in discrete oscillatory bursts rather than as sustained activity. We relate dynamic network activity as well as key synaptic characteristics to electrophysiological measurements. Our findings support the hypothesis that fast Hebbian short-term potentiation is a key WM mechanism.SIGNIFICANCE STATEMENT Working memory (WM) is a key component of cognition. Hypotheses about the neural mechanism behind WM are currently under revision. Reflecting recent findings of fast Hebbian synaptic plasticity in cortex, we test whether a cortical spiking neural network model with such a mechanism can learn a multi-item WM task (word list learning). We show that our model can reproduce human cognitive phenomena and achieve comparable memory performance in both free and cued recall while being simultaneously compatible with experimental data on structure, connectivity, and neurophysiology of the underlying cortical tissue. These findings are directly relevant to the ongoing paradigm shift in the WM field.

Between persistently active and activity‐silent frameworks: novel vistas on the cellular basis of working memory

“…Indeed, models show that a mixture of a rate code and STSP can maintain memories. For instance, Fiebig and Lansner proposed a model of WM that uses a combination of rate coding and a Hebbian form of spike timing–dependent STSP. This network is able to hold multiple items in memory at the same time.…”

Example Based Hebbian Learning may be sufficient to support Human Intelligence

“…An important question for the EBH hypothesis is: what is the structure of this conscious control process, and how can it be described in a way that is biologically plausible and useful as a component of a computational model? We do not speculate on the answer to this question here, but note that this process is closely related to phenomena studied under many labels, including conscious control (Schneider and Shiffrin, 1977;Shepherd, 2015), imagination (Ayman-Nolley, 1992;Asma, 2017), working memory (Fiebig and Lansner, 2017;Velichkovsky, 2017;Masse et al, 2019), and consciousness (Dehaene and Naccache, 2001;Tallon-Baudry, 2011;Koch et al, 2016;Tononi et al, 2016;Velichkovsky, 2017), and prior work in these areas may provide clues as to how this process emerges and operates, and how it may contribute mechanistically to EBH learning.…”

Probabilistic associative learning suffices for learning the temporal structure of multiple sequences

“…Here we have followed the modelling philosophy aimed at distilling the architecture of the network to its essential characteristics that support and control the phenomenon of interest (sequence learning). In the previous models of particular relevance to our work, complex spike based dynamics and rich biological detail were promoted to provide insights into the biophysical underpinnings of sequence learning in the cortex (Tully et al, 2016) and as a model of memory consolidation (Fiebig and Lansner, 2017). While the aforementioned contributions provide a more direct mapping between biology and the network, our approach, which reduces the network to its essential characteristics, necessarily dilutes that mapping.…”