A working memory model based on recurrent neural networks using reinforcement learning.

Area of Science:

• Computational neuroscience
• Cognitive neuroscience
• Machine learning

Background:

• Prefrontal cortex (PFC) neurons exhibit dynamic single-unit activity yet stable population coding for working memory.
• The neural computation mechanisms underlying this PFC activity in working memory remain unclear.
• Understanding neural network dynamics is crucial for deciphering working memory processes.

Purpose of the Study:

• To explore the neural computation mechanism of working memory using a novel computational approach.
• To simulate a spatial working memory task using a recurrent neural network (RNN) model.
• To investigate how RNN activity patterns relate to prefrontal cortex neuron activity.

Main Methods:

• Trained a recurrent neural network model with a decision and a baseline network using reinforcement learning.
• The model learned a spatial working memory task, involving stimulus information maintenance.
• Analyzed unit and population activity dynamics, including temporal dynamics and low-dimensional encoding.

Main Results:

• The RNN model successfully performed the spatial working memory task.
• Model unit activity showed temporal dynamics and preferred direction selectivity, mirroring PFC neuron activity.
• Population activity stably encoded stimulus information in a low-dimensional subspace, improving with learning.

Conclusions:

• The developed RNN model provides a viable simulation for spatial working memory tasks.
• The model offers insights into how PFC neurons achieve stable information representation through population dynamics.
• This approach enhances understanding of neural computation in working memory and PFC function.