Hebbian Learning

“Cells that fire together wire together” — synapses strengthen when pre- and post-synaptic activity coincide.

Introduced by Donald Hebb in The Organization of Behavior (1949). Underlies the naming of memory tools like Hebbian Vault.

Biological mechanism

  • LTP (Long-Term Potentiation) — sustained synaptic strengthening after high-frequency stimulation; first reported by Bliss & Lømo (1973) in rabbit hippocampus.
  • STDP (Spike-Timing-Dependent Plasticity) — a presynaptic spike ~10 ms before a postsynaptic spike produces potentiation; reverse order produces depression. Requires NMDA-type glutamate receptors and a postsynaptic Ca²⁺ rise.
  • Synaptic tagging & capture — transient “tags” at activated synapses capture plasticity-related proteins, converting early to late LTP.
  • Neuromodulation — dopamine, acetylcholine, norepinephrine gate plasticity by shifting LTP/LTD thresholds (context-dependent learning).
  • Homeostasis — synaptic scaling acts as negative feedback to prevent runaway potentiation.

Formal rules

RuleFormProperty
Basic HebbΔw ∝ η·x·yCorrelational, but unstable without normalization
Oja’s ruleẇ = ⟨yx⟩ − ⟨y²⟩wWeight normalization → principal-component extraction
BCM rulesliding threshold θ_MPotentiate if y > θ_M, else depress; stable metaplasticity
STDPexponential timing windowsSpiking-network online implementation via pre/post traces

Computational models

  • Associative memory — Hopfield networks via outer-product storage W = Σ xᵘ(xᵘ)ᵀ (pattern completion)
  • ICA / source separation — (anti-)Hebbian rules maximize non-Gaussianity
  • Sparse coding — anti-Hebbian decorrelation produces localized receptive fields
  • Self-organizing maps — Hebbian-like updates with neighborhood functions preserve topography

AI applications

  • Engram Neural Network (ENN) — stores an outer-product Hebbian memory trace; reports comparable performance to RNN baselines on MNIST/CIFAR-10 with faster training.
  • Neuromorphic hardware — BrainScaleS implements Hebbian/STDP in analog/digital hardware with large speedups over biological real-time.
  • Open-source — PyTorch-Hebbian libraries enable reproducible Hebbian training of conv layers.

vs. gradient-based learning

Hebbian/local plasticity is biologically plausible and intrinsically local, but struggles to propagate supervised teacher signals through deep hierarchies (deep credit assignment). Hybrid schemes — e.g. Random Feedback Local Online (RFLO), meta-learned plasticity — combine local eligibility traces, feedback projections, and modulatory signals to approximate structured credit assignment while retaining locality.

Open questions

  • No systematic large-scale benchmarks of Hebbian vs. gradient-based deep learning on modern tasks.
  • How complex neuromodulatory states implement task-level gating across cortical hierarchies remains uncharacterized.

Researched via Tavily tvly research, 2026-05-30.