Neural Mechanisms of Working Memory Accuracy Revealed by Recurrent Neural Networks.

Yuanqi Xie,Xin Zhou,Yichen Henry Liu,Christos Constantinidis

doi:10.3389/fnsys.2022.760864

Yuanqi Xie, Xin Zhou + Show 2 more

Open Access

https://doi.org/10.3389/fnsys.2022.760864

Copy DOI

Abstract

Understanding the neural mechanisms of working memory has been a long-standing Neuroscience goal. Bump attractor models have been used to simulate persistent activity generated in the prefrontal cortex during working memory tasks and to study the relationship between activity and behavior. How realistic the assumptions of these models are has been a matter of debate. Here, we relied on an alternative strategy to gain insights into the computational principles behind the generation of persistent activity and on whether current models capture some universal computational principles. We trained Recurrent Neural Networks (RNNs) to perform spatial working memory tasks and examined what aspects of RNN activity accounted for working memory performance. Furthermore, we compared activity in fully trained networks and immature networks, achieving only imperfect performance. We thus examined the relationship between the trial-to-trial variability of responses simulated by the network and different aspects of unit activity as a way of identifying the critical parameters of memory maintenance. Properties that spontaneously emerged in the artificial network strongly resembled persistent activity of prefrontal neurons. Most importantly, these included drift of network activity during the course of a trial that was causal to the behavior of the network. As a consequence, delay period firing rate and behavior were positively correlated, in strong analogy to experimental results from the prefrontal cortex. These findings reveal that delay period activity is computationally efficient in maintaining working memory, as evidenced by unbiased optimization of parameters in artificial neural networks, oblivious to the properties of prefrontal neurons.

Highlights

Working memory, the ability to maintain information in mind over a period of seconds is a core cognitive function, essential for higher human faculties (Baddeley, 2012)
Delay period firing rate and behavior were positively correlated, in strong analogy to experimental results from the prefrontal cortex. These findings reveal that delay period activity is computationally efficient in maintaining working memory, as evidenced by unbiased optimization of parameters in artificial neural networks, oblivious to the properties of prefrontal neurons
We trained leaky Recurrent Neural Networks (RNNs) to perform multiple working memory tasks: the Oculomotor Delayed Response or ODR task (Figure 1A); a variant of the ODR task requiring the subject to remember the location of a cue and ignore the presentation of a subsequent distractor (ODRD task in Figure 1B); and response inhibition tasks, as we have described recently (Liu et al, 2021)

Summary

Introduction

The ability to maintain information in mind over a period of seconds is a core cognitive function, essential for higher human faculties (Baddeley, 2012). Generation of persistent activity has been modeled as a continuous attractor by biophysically inspired network models that generate a bump (peak) of activity representing the stimulus to be remembered (Compte et al, 2000) Predictions of these models about how neuronal activity, variability, and correlation and how these relate to performance of working memory tasks are borne by neurophysiological data (Wimmer et al, 2014; Barbosa et al, 2020). It is unclear, whether underlying assumptions of bump attractor models are realistic and whether their simplified structure is truly compatible with the diversity and variability of real neuronal responses. Many empirical results observed in neurophysiological recordings are often difficult to interpret in the context of the bump attractor (Qi et al, 2021)

Methods

Results

Discussion

Conclusion