Interactive figures for teaching how synapses change

Neural Plasticity

Learning, in a nervous system, is synapses changing strength. These figures take the rules that describe that change and let you drive them: set the interval between two spikes and watch a synapse potentiate or depress, train patterns into a network until it can recall them from a fragment, and push the cell that all of it happens to until it stops doing what the textbook says. Every figure runs offline, adapts to the projector, and states where its own model stops being true.

Learning rules

Two answers to the same question: what, exactly, does a synapse measure before it decides to change?

Spike-Timing-Dependent Plasticity Set the interval between the pre- and post-synaptic spike and run a pairing protocol. A few milliseconds either side of coincidence decides whether the synapse is strengthened or weakened, and sweeping the interval reproduces the measured learning window. LTP & LTD · pairing protocols · weight bounds Hebbian Cell Assemblies Train two patterns into a recurrent network and watch the weight matrix separate into blocks, then recall a whole assembly from two cells of it. The payoff Hebb was after: a memory addressed by its contents. Pattern completion · synaptic scaling · weight matrices

The cell it happens to

Plasticity rules are written in terms of spikes. This is where the spikes come from.

Stochastic Leaky Integrate-and-Fire The smallest model that still behaves like a neuron: a leaky capacitor with a threshold. Membrane trace, raster, firing rate and the f–I curve against its own closed-form prediction, with noise strong enough to make the cell fire below rheobase. f–I curve · rheobase · ISI statistics

How the three fit together

Read them in the order that a signal travels. The integrate-and-fire cell turns current into spike times, and shows that when the input is noisy those times are irregular rather than clock-like. The STDP window takes two of those spike times, one on each side of a synapse, and turns their difference into a change in one weight. The assembly figure applies a rule of that kind across a whole network, many times, and shows what falls out: structure that no individual synapse could have known about.

That last step is the claim worth being sceptical about, and each figure is built so you can attack it. The learning rules here are local, in the strict sense that a synapse only ever sees its own two cells. Whether local rules are enough to build a memory is an empirical question, not a theorem, and every one of these models has a section saying where it is known to break.