A neural network for visual working memory that accounts for memory binding errors

Abstract

Binding errors, also called swap errors, occur in working memory (WM) tasks when the participant fails to report the feature of a previously presented target but the response is accurate relative to a non-target stimulus [1]. These errors reflect the failure of the system to maintain bundled through memory the conjunction of features that define one object. The brain mechanisms that maintain integrated several features of an item in one memory remain unknown. We explore the hypothesis that synchrony of different neural assemblies coding each for a feature of an item plays the main role [2]. To test the synchrony hypothesis, we built a network model for the storage of multiple items defined by one color and one location in WM. The model is composed of two networks for WM [3], one representing colors and the other one locations. These two networks are then connected via weak cortico-cortical excitation. With this model we are able to maintain persistent memory bumps that encode colors and locations in each respective network. Fast recurrent excitation within each network induced γ oscillations during bump activity through the interplay of fast excitation and slower feedback inhibition [3]. Spectral power of network activity in the γ range of frequencies increased with the number of stored items, as it has been reported both in humans and monkey studies [4]. Binding between features was effectively accomplished through the synchronization of γ oscillations between bumps across the two networks. As a result, different memorized items were held at different phases of a global network oscillation, whose frequency increased with WM load. In some simulations swap errors arose: color bumps abruptly changed their phase relationship with location bump. Furthermore, by systematically decreasing the distance between different memory items, thus increasing potential interference within trials [5], the model predicts that swap errors should increase.