Reproduce “Advancing the Biological Plausibility and Efficacy of Hebbian Convolutional Neural Networks” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Julian Jimenez Nimmo, Esther Mondragon
Researchers at City St George’s, University of London developed and optimized a biologically plausible Hebbian Convolutional Neural Network using Hard Winner-Takes-All (Hard-WTA) competition, setting a new state-of-the-art for this approach. The Optimal-HardWTA model achieved 76% accuracy on CIFAR-10, statistically matching backpropagation in shallow networks, and outperformed backpropagation by over 11% accuracy on the more complex STL-10 dataset.
Key insights
- Optimized Hard-WTA competition, when combined with biologically inspired mechanisms, can achieve performance competitive with or superior to backpropagation in certain image classification tasks, even in deeper networks.
- Incorporating specific biological principles like lateral inhibition and the BCM learning rule is crucial for both enhancing the biological plausibility and improving the efficacy of Hebbian CNNs.
- Hebbian learning, particularly with Hard-WTA, leads to more interpretable feature representations, characterized by clearer class clustering in UMAP embeddings and sparse, biologically analogous weight distributions.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.