A recurrent neural network model for a decision-making task based on sequential evidence accumulation.

Area of Science:

• Neuroscience
• Computational Neuroscience
• Cognitive Science

Background:

• Decision-making relies on accumulating evidence, a process involving neural activity across brain regions.
• The dorsolateral prefrontal cortex (DLPFC) is known to play a role in evidence accumulation for decisions.
• Existing computational models lack clarity on dynamic integration across timescales and stability maintenance.

Purpose of the Study:

• To propose and investigate a recurrent neural network (RNN) model with reinforcement learning for understanding neural computations in evidence accumulation.
• To explore how serial information is dynamically integrated across multiple timescales.
• To elucidate mechanisms for maintaining the stability of accumulated evidence states.

Main Methods:

• Developed a recurrent neural network (RNN) model incorporating reinforcement learning.
• Simulated the model's performance on an evidence accumulation decision-making task.
• Analyzed population activity patterns of recurrent units and network dynamics.

Main Results:

• The RNN model successfully learned and performed the evidence accumulation task.
• Distinct neural coding patterns were observed: transient responses to instantaneous evidence and gradual accumulation reflecting integrated evidence.
• The network demonstrated dynamic integration of evidence over time, not just tracking instantaneous inputs.
• Disrupting network units or connections impaired decision-making performance, validating the model's architecture.

Conclusions:

• The proposed RNN model effectively simulates sequential evidence accumulation for decision-making.
• The model provides a computational framework for understanding neural-level evidence accumulation.
• Findings suggest the critical role of network architecture in dynamic evidence integration and decision stability.