Engineering recurrent neural networks from task-relevant manifolds and dynamics.

Area of Science:

• Computational neuroscience
• Cognitive science
• Machine learning

Background:

• Cognitive processes rely on transformations of neural population representations.
• Recurrent neural networks (RNNs) model these processes, but understanding their connectivity and dynamics is challenging.
• Existing models lack interpretability in how network structure generates task-solving dynamics.

Purpose of the Study:

• To present a novel method for synthesizing RNN connectivity with specified, interpretable dynamics.
• To apply this method to model a working memory task using a drift-diffusion process.
• To explore input-driven control of network dynamics for cognitive flexibility and representation geometry's impact on capacity.

Main Methods:

• Developed a method to derive RNN connectivity based on desired task dynamics.
• Synthesized a network implementing a drift-diffusion process on a ring manifold for working memory.
• Investigated how external inputs can modulate network dynamics and analyzed representation geometry.

Main Results:

• Successfully synthesized RNNs with interpretable, task-specific dynamics.
• Demonstrated a working memory model based on a drift-diffusion process on a manifold.
• Showcased input-driven control for cognitive flexibility and explored capacity-representation geometry relationships.

Conclusions:

• The proposed method enables the creation of interpretable neural network models for cognitive tasks.
• Understanding neural computation through low-dimensional manifold dynamics is facilitated by this approach.
• This work provides a framework for designing neural networks with specific computational properties.