A novel framework for expanding RNNs with biophysical detail to solve cognitive tasks

Area of Science:

• Computational neuroscience
• Neural networks
• Cognitive modeling

Background:

• Recurrent neural networks (RNNs) excel at cognitive tasks but use abstract neurons.
• Biophysically realistic neurons offer computational benefits but pose training challenges.
• Bridging abstract and biophysical models is crucial for understanding neural circuits.

Purpose of the Study:

• Develop a biophysically detailed reservoir computing (BRC) framework.
• Extract mechanistic insights from biophysical neural models for cognitive tasks.
• Investigate the impact of biophysical properties on working memory.

Main Methods:

• Constructed a BRC network with multicompartment excitatory neurons and active dendrites.
• Trained the BRC network on a simplified working memory task.
• Analyzed the effects of input type (AMPA vs. NMDA) and location (dendrite vs. soma).

Main Results:

• NMDA receptor inputs were significantly more efficient for solving the working memory task.
• Dendritic NMDA receptor inputs were uniquely effective, while dendritic AMPA inputs failed.
• Distinct network dynamics and computational principles emerged based on input type and location.

Conclusions:

• BRC framework facilitates training of biophysically detailed neural models.
• Dendritic NMDA receptors play a critical role in working memory tasks.
• This approach provides a foundation for understanding neural mechanisms underlying cognition.