Hebbian crosstalk and input segregation.

Area of Science:

• Computational Neuroscience
• Machine Learning Theory
• Neural Development

Background:

• Hebbian synapses learn based on input/output correlations, reflecting input statistical structure.
• Non-connection-specific strength adjustments, termed 'crosstalk', can distort or inhibit neural learning.
• Crosstalk in neural learning is analogous to mutation errors in genetic evolution.

Purpose of the Study:

• To analyze the impact of crosstalk on Hebbian learning, specifically within the Oja model.
• To investigate how crosstalk affects ocular segregation, a process requiring negative input correlations.
• To determine if crosstalk can induce bifurcations in learning outcomes, analogous to 'error catastrophes' in genetics.

Main Methods:

• Analysis of Hebbian learning driven by pairwise input correlations using the classical Oja model.
• Mathematical modeling to examine the effect of crosstalk on ocular segregation under varying input correlations.
• Investigation of critical states and behavior near bifurcations for statistically unbiased and weakly biased inputs.

Main Results:

• Crosstalk leads to learning the principal eigenvector of EC (input covariance matrix times error matrix).
• For statistically unbiased inputs, crosstalk induces a bifurcation from segregating to non-segregating outcomes at a critical correlation value.
• Even low levels of crosstalk observed in the brain could potentially impede normal neural development.

Conclusions:

• Crosstalk can induce bifurcations in simple Hebbian learning models under specific conditions.
• The findings suggest that minimal crosstalk is crucial for successful neural learning and development.
• While complex learning may mitigate these effects, the Oja model highlights potential risks of crosstalk in neural systems.