Scholarlink

<p>Decision making typically entails the accumulation of evidence, a process supported by coordinated neural activity across multiple brain regions. Numerous studies have shown that the dorsolateral prefrontal cortex (DLPFC) participates in evidence accumulation for decision-making. Single neurons in the DLPFC display ramping responses related to accumulated evidence, and neuronal populations stably encode the integrated evidence that supports the transformation from sensory inputs to actions. Previous studies have proposed computational models of evidence accumulation in decision-making. However, it is still unclear how serial information can be dynamically integrated across multiple timescales and how the stability of accumulated states is maintained. To address these issues, we proposed a recurrent neural network (RNN) model with reinforcement learning to probe the neural computations underlying evidence accumulation in the decision-making task. Simulation results show that the model successfully learned to perform the evidence accumulation decision-making task. The population activity of recurrent units shows distinct coding patterns: one subset displayed transient responses to instantaneous evidence under different stimulus conditions; another subset exhibited activity that gradually increased or decreased with the amount of accumulated evidence in favor of the preferred target, thereby reflecting the process of evidence accumulation. Further analyses indicated that the network did not simply track instantaneous evidence, but rather tended to integrate these signals over time to complete the decision. This property resembles the behavior of DLPFC neurons in similar tasks, and highlights the model’s capacity for dynamic integration of evidence. Furthermore, we found that both suppressing the activity of specific units and disrupting network connections impaired the model’s decision-making performance, thereby validating the critical role of this network architecture in executing the task. Taken together, the simulated results suggest that the model accumulates sequentially inputted evidence for decision-making and offers a possible computational way to understand how evidence is accumulated in the neural level.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

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

  • 期刊论文

  • 发表日期：

    2026年6月30日

查看全文
  • Jiaping Liu,
  • Yihong Wang,
  • Xuying Xu,
  • Xiaochuan Pan

摘要

Decision making typically entails the accumulation of evidence, a process supported by coordinated neural activity across multiple brain regions. Numerous studies have shown that the dorsolateral prefrontal cortex (DLPFC) participates in evidence accumulation for decision-making. Single neurons in the DLPFC display ramping responses related to accumulated evidence, and neuronal populations stably encode the integrated evidence that supports the transformation from sensory inputs to actions. Previous studies have proposed computational models of evidence accumulation in decision-making. However, it is still unclear how serial information can be dynamically integrated across multiple timescales and how the stability of accumulated states is maintained. To address these issues, we proposed a recurrent neural network (RNN) model with reinforcement learning to probe the neural computations underlying evidence accumulation in the decision-making task. Simulation results show that the model successfully learned to perform the evidence accumulation decision-making task. The population activity of recurrent units shows distinct coding patterns: one subset displayed transient responses to instantaneous evidence under different stimulus conditions; another subset exhibited activity that gradually increased or decreased with the amount of accumulated evidence in favor of the preferred target, thereby reflecting the process of evidence accumulation. Further analyses indicated that the network did not simply track instantaneous evidence, but rather tended to integrate these signals over time to complete the decision. This property resembles the behavior of DLPFC neurons in similar tasks, and highlights the model’s capacity for dynamic integration of evidence. Furthermore, we found that both suppressing the activity of specific units and disrupting network connections impaired the model’s decision-making performance, thereby validating the critical role of this network architecture in executing the task. Taken together, the simulated results suggest that the model accumulates sequentially inputted evidence for decision-making and offers a possible computational way to understand how evidence is accumulated in the neural level.