Beyond the Delay Neural Dynamics: a Decoding Strategy for Working Memory Error Reduction

Abstract Understanding how the brain preserves information despite intrinsic noise is a fundamental question in working memory. A typical delayed-response task consists of a delay epoch for maintaining information, go and response epochs (decoding phase) for decoding maintained neural state to an output action. While previous works have focused on the delay neural dynamics, as another key phase in working memory, the role of decoding phase in memory error reduction has not been investigated: what and how the maintained neural state is decoded to an output action; and how this decoding process can help reducing the memory error? We address these questions by training artificial recurrent neural networks (RNNs) to execute a color delayed-response task. We found the trained RNNs learned to reduce the memory error of the high-probability-occur colors (common colors) by decoding/attributing a broader range of neural state space to common colors. This decoding strategy can be further explained by both the converging neural dynamics and a non-dynamic, biased readout process during the decoding phase. Our findings provide testable prediction of the critical role of the decoding phase in memory processes, suggesting that neural systems deploy multiple strategies across different phases to reduce the memory errors.Significance Statement Preserving information under noise is of crucial in working memory. A typical delayed-response experiment consists of a delay epoch for maintaining information, and a go and response epoch (decoding phase) for decoding the maintained neural state into output information. While the delay neural dynamics has 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 neural state to that color. This strategy is partially supported by a non-dynamic readout process. Our results suggest that neural networks can utilize diverse strategies, beyond delay neural dynamics, to reduce memory errors..

Medienart:	Preprint

Erscheinungsjahr:	2024
Erschienen:	2024

Enthalten in:	bioRxiv.org - (2024) vom: 13. März Zur Gesamtaufnahme - year:2024

Sprache:	Englisch

Beteiligte Personen:	Ye, Zeyuan [VerfasserIn] Li, Haoran [VerfasserIn] Tian, Liang [VerfasserIn] Zhou, Changsong [VerfasserIn]

Links:	Volltext [kostenfrei]

Themen:	570 Biology

doi:	10.1101/2022.06.01.494426

funding:
Förderinstitution / Projekttitel:

PPN (Katalog-ID):	XBI036183245