Sudden restructuring of memory representations in recurrent neural networks with repeated stimulus presentations

Area of Science:

• Computational Neuroscience
• Cognitive Science
• Machine Learning

Background:

• Human learning curves often show discontinuous performance jumps, unlike smooth group-averaged curves.
• Nonlinear systems exhibit thresholding effects, which can be modeled abstractly.
• Associative learning mechanisms in neural networks are key to understanding cognitive processes.

Purpose of the Study:

• To investigate discontinuous changes in learning performance using computational models.
• To explore how Amari-Hopfield networks with Hebbian learning rules exhibit learning.
• To compare model-generated learning patterns with observed human learning behavior.

Main Methods:

• Simulating Amari-Hopfield networks with Hebbian learning rules.
• Repeatedly exposing the network to a single stimulus.
• Analyzing attractor basin size changes and transition graph structures.

Main Results:

• Attractor basin size for a target stimulus increased in discrete jumps, not gradual changes.
• The distribution of performance jumps followed a lognormal distribution (heavy-tailed).
• Newly acquired states formed hierarchical tree structures with branch sizes following a power law distribution.

Conclusions:

• Simple nonlinear network models exhibit discontinuous learning performance, mirroring human behavior.
• These findings suggest potential network mechanisms for learning through repeated exposure.
• Future research can explore these mechanisms in more biologically detailed models.