Unsupervised Hebbian Learning from an Artificial Intelligence Perspectives (www.mdpi.com)

This paper revisits Hebbian and anti-Hebbian learning through an AI lens, systematically defining mathematical update rules, activation functions, and practical implementations, then testing them on vision benchmarks (Fashion-MNIST, KMNIST, MNIST) with a set of explicit data-fusion protocols (s0–s7) including sequential training, independent training plus merging, and combined-dataset training. Key empirical findings show that Hebbian-style synapses can learn useful representations but are prone to overfitting (visible in degraded performance and characteristic synapse visualizations), that activation choice strongly shapes the learned features, and that data-fusion via synaptic weight merging (an explicit equation and a hybrid parameter) materially affects downstream performance—paralleling recent “model soup” observations. Technically, the work ties classical biologically plausible rules (Hebb/anti-Hebb, winner-take-all motifs) to modern ML recipes: experiments use batch size 100, hidden layer size matching input dimensionality, tunable power and anti-Hebbian strength/indices, numerical stability thresholds, Adam optimizer and cross-entropy in supervised comparisons, and learning-rate schedules (linear decay or cosine annealing from 0.01 → 0 over 200 epochs). Implications: Hebbian schemes are viable lightweight pretraining or representation-learning primitives, but require careful regularization, activation choices, and principled merging strategies when leveraging multiple unlabeled datasets to avoid catastrophic specialization.