Hebbian errors in learning: an analysis using the Oja model | JoVE Visualize

Area of Science:

Computational Neuroscience
Machine Learning
Synaptic Plasticity

Background:

Hebbian learning, a key mechanism for synaptic plasticity, is assumed to be synapse-specific.
Recent findings suggest Hebbian learning may not be perfectly specific, possibly due to factors like calcium diffusion.
This lack of specificity is proposed as analogous to mutations in evolutionary processes.

Purpose of the Study:

To quantitatively examine the impact of Hebbian learning inspecificity on unsupervised learning models.
To extend the Oja unsupervised learning model to incorporate Hebbian inspecificity.
To analyze the relationship between learning accuracy and parameters reflecting synapse density and connectivity.

Main Methods:

Developed a modified Oja unsupervised learning model incorporating an error matrix (E) to represent crosstalk between synaptic connections.
Analyzed the convergence properties of the modified algorithm, showing it converges to the leading eigenvector of EC, where C is the input covariance matrix.
Investigated the dependence of learning accuracy on synaptic parameter (b), number of inputs (n), and input correlation through analytical and computational methods.

Main Results:

The modified algorithm converges to the leading eigenvector of EC, deviating from the classical PC1 result when inspecificity is present.
Learning accuracy is found to increase with the synaptic parameter (b), which reflects synapse density and diffusion, particularly at biologically realistic correlation strengths.
Increased Hebbian inspecificity, especially with weak input correlations, significantly degrades learning performance, though some learning persists.

Conclusions:

Hebbian learning mechanisms in the brain, when lacking specificity, can lead to suboptimal or degraded learning outcomes.
The proposed model quantitatively links Hebbian crosstalk to reduced learning accuracy, particularly in scenarios with weak input correlations.
Synaptic crosstalk, arising from dense synaptic connectivity, is an inherent factor that degrades the effectiveness of Hebbian unsupervised learning in neural networks.