Representational drift as a result of implicit regularization.

Area of Science:

• Neuroscience
• Artificial Intelligence
• Machine Learning

Background:

• Neuronal tuning changes over time in constant environments, a phenomenon known as representational drift.
• This drift is hypothesized to result from continuous learning under noisy conditions, but its mechanisms require further investigation.

Purpose of the Study:

• To investigate the underlying mechanisms of representational drift in neural networks.
• To analyze the temporal dynamics of neuronal activity during prolonged learning in a stable environment.

Main Methods:

• Trained an artificial neural network on a simplified navigational task.
• Analyzed four independent datasets of CA1 neuronal activity from different laboratories.
• Examined changes in neuronal activity sparseness and spatial information over time.

Main Results:

• The artificial network rapidly achieved high performance, with units showing spatial tuning.
• Continued training led to activity sparsification, occurring much slower than initial learning.
• CA1 neurons in real brains also exhibited increased sparseness and spatial informativeness with prolonged environmental exposure.

Conclusions:

• Learning can be characterized by three overlapping phases: fast familiarity, slow implicit regularization, and a steady state of null drift.
• Representational drift dynamics offer insights into the underlying learning algorithms.
• The findings suggest a unified mechanism for representational drift across artificial and biological neural networks.