Standard deep learning adjusts weights globally — a single error signal propagates
backward through the entire network via backpropagation. Hebbian learning
is different: weights update locally, based only on what two connected neurons are doing
right now. No global error. No backward pass. Just the conversation between two nodes.
The rule: Δwᵢ = α · xᵢ · y
— the weight change between input node xᵢ and output node y equals the learning rate α
times the product of their activations. If both fire strongly and simultaneously,
the connection grows. If one fires and the other doesn't, the weight may weaken.
The connection is a record of co-activation.
Color collision. Each layer carries a spectral hue — coral at the input,
moving through orange, green, blue, violet at the output. When a signal from layer A
reaches layer B, the arriving hue and the destination hue mix additively at the collision point.
The resulting color is the blended activation — a visual encoding of
Δw at that instant. Connections that repeat the same path grow brighter in their blended color.
Ricochets happen when a signal bounces between two high-weight connections
before resolving forward — a visual analogue of resonance in a recurrent net.
In real Hebbian networks, strong bidirectional weights can create sustained oscillation,
which underlies associative memory and pattern completion.
Weight history accumulates across firings. After many signals,
the connections that reliably co-activate glow brightest — the network has
learned a pattern without ever being told what the pattern should be.
This is the basis of Hopfield networks, early perceptrons, and modern sparse coding.
click any input node to target it · watch Δw update per connection · weights accumulate across firings