Benchmarking local Hebbian learning rules for memory storage and prototype extraction
New benchmark of seven Hebbian rules reveals Bayesian approaches dominate memory storage and recall
Associative memory, or content-addressable memory, is a core function in both computer science and cognitive neuroscience. This paper by Lansner et al. (arXiv:2605.01074) systematically benchmarks seven different Hebbian learning rules in non-modular and modular recurrent networks using winner-take-all (WTA) dynamics. The study focuses on two key capabilities: pattern storage (memory capacity and weight information) and prototype extraction, where the network must recall the original prototype from a distorted instance. The rules are tested on moderately sparse binary patterns, and the benchmark also evaluates sensitivity to correlations in the training data.
Results show a clear hierarchy: the classic additive Hebb rule has the worst capacity, covariance learning is robust but offers only moderate performance, and Bayesian-Hebbian learning rules consistently deliver the highest capacity across almost all test conditions. These findings provide a rigorous comparison for researchers designing neuromorphic memory systems and reinforcement learning architectures that rely on local, biologically plausible learning rules. The Bayesian-Hebbian approach, which incorporates prior knowledge about pattern statistics, emerges as the most effective for both storage and prototype extraction tasks.
- Seven Hebbian learning rules benchmarked: additive, covariance, and Bayesian-Hebbian variants in WTA networks
- Bayesian-Hebbian rules achieved highest memory storage and prototype extraction capacity under most conditions
- Additive Hebb rule performed worst; covariance learning was robust but only moderate capacity
Why It Matters
Better Hebbian learning rules could enable more efficient neuromorphic memory systems and biologically inspired AI models