One Step Back, Two Steps Forward: Interference and Learning in Recurrent Neural Networks.

Area of Science:

• Computational Neuroscience
• Machine Learning
• Cognitive Science

Background:

• Artificial neural networks (ANNs) are increasingly used to model neural recordings from animals during cognitive tasks.
• The evolution of network dynamics during the learning process, however, remains largely unexplored in ANNs.
• Experimental studies often focus on trained animals, neglecting neural activity throughout the learning phase.

Discussion:

• This study analyzes ANNs trained on memory and pattern generation tasks, linking task functions to dynamical attractors (line attractors, limit cycles).
• The sequential, trial-by-trial nature of learning significantly impacts learning trajectories and outcomes.
• Least Mean Squares (LMS), a biologically plausible learning rule, suffers from trial interference (forgetting), slowing convergence compared to algorithms like FORCE.

Key Insights:

• Forgetting, manifested as the destruction of previously learned dynamical objects, significantly obstructs LMS learning.
• The degree of interference depends on the correlation between trials.
• Specific components of the FORCE algorithm mitigate trial interference, leading to faster convergence.

Outlook:

• Insights into trial interference can guide the design of experimental protocols to minimize this effect in biological systems.
• Observing neural activity and behavior throughout learning may help infer underlying learning rules.
• Understanding learning-rule-dependent interference is crucial for developing more efficient artificial and biological learning models.