Learning without backprop

The artificial neuron of deep learning takes a number and returns a number. A neuron in the brain lives in time: the voltage across its membrane slowly fills with incoming current, drains through a leak, and the moment it crosses a threshold it fires a brief spike and resets. Information travels in the timing of spikes, not in the size of numbers. aura-mind is built from neurons like these.

Current below threshold never produces a single spike however long it lasts: the leak drains it all away. Past the threshold the rate climbs quickly, then saturates because of the short rest after every spike (the refractory period). A little noise makes a neuron just under threshold fire now and then; the brain uses that randomness as a feature.

Backprop is a global computation that carries the error at the end of the network back through every layer. The brain does not seem to have such a channel. Instead every synapse sees only what happens at its own two ends. The STDP rule says: if the sending neuron fired just before the receiver, so it may have caused it, the connection strengthens; if it fired just after, it weakens.

The rule is entirely local: a synapse needs to know nothing about the rest of the network to learn. aura-mind adds a third factor, a modulatory signal like reward or surprise, so that “we fired together” only sticks if it turned out to be useful.

Can local learning rules build a memory? A classic example: a network where every neuron connects to every other. To store a pattern you update each connection using only whether its two neurons are active together, minus the average activity: Hebb’s covariance rule. Then you corrupt the pattern and let the network go; each neuron looks at the total input from its neighbours, a shared inhibitory pool silences all but the most strongly driven, and step by step the network slides back into the stored pattern.

With a few patterns the network repairs even heavy damage. Keep adding: past a point recall collapses and the network falls into ghosts that blend patterns together. For letters that resemble each other, like E and Z, the collapse comes very early; store all eight letters and try corrupting E. Random patterns fit more, but not without limit either. This is a small model of catastrophic forgetting: new knowledge writes over the old. All of aura-mind’s research is about finding an architecture that prevents that collapse without giving up locality.