A Unified Dynamic and Computational Mechanism for Persistent and Transient Neural Activity Patterns During Delayed-Response Tasks

Abstract

Understanding how the brain preserves information despite intrinsic noise is a fundamental question in working memory. Typical working memory tasks consist of delay phase for maintaining information, and decoding phase for retrieving information. While previous works have focused on the delay neural dynamics, it is poorly understood whether and how the neural process during decoding phase reduces memory error. We studied this question by training recurrent neural networks (RNNs) on a color delayed-response task. We found that the trained RNNs reduce the memory error of high-probability-occurring colors (common colors) by decoding/attributing a broader range of neural states to them during decoding phase. This decoding strategy can be further explained by a continuing converging neural dynamics following delay phase and a non-dynamic biased readout process. Our findings highlight the role of the decoding phase in working memory, suggesting that neural systems deploy multiple strategies across different phases to reduce memory errors. SignificancePreserving information under noise is crucial in working memory. A typical working memory task consists of a delay phase for maintaining information, and a decoding phase for decoding the maintained into an output action. While the delay neural dynamics have been intensively studied, the impact of the decoding phase on memory error reduction remains unexplored. We trained recurrent neural networks (RNNs) on a color delayed-response task and found that RNNs reduce memory error of a color by decoding a larger portion of the neural state to that color. This strategy is supported both by a converging neural dynamic, and a non-dynamic readout process. Our results suggest that neural networks can utilize diverse strategies, beyond delay neural dynamics, to reduce memory errors.