A Robust Model of Gated Working Memory

Abstract

Gated working memory is defined as the capacity of holding arbitrary information at any time in order to be used at a later time. Based on electrophysiological recordings, several computational models have tackled the problem using dedicated and explicit mechanisms. We propose instead to consider an implicit mechanism based on a random recurrent neural network. We introduce a robust yet simple reservoir model of gated working memory with instantaneous updates. The model is able to store an arbitrary real value at random time over an extended period of time. The dynamics of the model is a line attractor that learns to exploit reentry and a nonlinearity during the training phase using only a few representative values. A deeper study of the model shows that there is actually a large range of hyper parameters for which the results hold (number of neurons, sparsity, global weight scaling, etc.) such that any large enough population, mixing excitatory and inhibitory neurons can quickly learn to realize such gated working memory. This suggests this property could be an implicit property of any random population, that can be acquired through learning. Furthermore, considering working memory to be a physically open but functionally closed system, we give account on some counter-intuitive electrophysiological recordings.

A journey in ESN and LSTM visualisations on a language task

“…Moreover, it could also help to understand the influence of the feedback when outputs are fed-back to the reservoir. Given that feeding back outputs into the reservoir can reduce the dimensionality of the itnernal states (especially for symbolic tasks) [52][53][54], studying this effect more deeply with RSSviz would provide more insights on the complex influence of the feedback in general. We believe that studying the combination of both unsupervised rules and feedback could lead to new learning rules for ESNs.…”

ReservoirPy: An Efficient and User-Friendly Library to Design Echo State Networks

“…Thus, online learning proposes a lighter approach to train reservoir with less computational demand while still achieving compatible level of accuracy. More importantly perhaps, online incremental learning methods are crucial for computational neuroscience models [14] and developmental experiments in cognitive science (developmental psychology, robotics, ...) [6,11]. Current implementation of online learning in ReservoirPy is based on FORCE learning method [15], and resides in a separate class: ESNOnline.…”

“…Several members of our team and students already used it for different tasks and purposes (e.g. to build Computational Neuroscience models and Human-Robot Interaction modules [6,11,14]), it is now time to share it more extensively.…”

Using Conceptors to Transfer Between Long-Term and Short-Term Memory

“…We consider the gating task described in [24] and two additional variants. In this task the model receives an input V that is continuously varying over time and another input being either 0 or 1 (trigger or gate T ).…”

“…Based on the original RC paradigm, we have shown in [24] how a reservoir with feedback connections can implement a gated working memory, i.e. a generic mechanism to maintain information at a given time (corresponding to when the gate is on).…”