Field Guide · models · the mind as a system
Memory as an attractor
Recall as flowing downhill, not looking up.
Click cells to corrupt the pattern, then hit recall — the network settles into the nearest stored memory.
Three patterns are stored in this little network — the letters B · S · H — not in any cell, but in the weights between all of them. Corrupt a pattern by clicking cells or adding noise, then hit recall: every neuron repeatedly flips to agree with the rest, and the whole state slides downhill until it lands in the nearest stored memory. Nothing is fetched from an address. The memory is the shape of the landscape.
This is a Hopfield network (1982), the model that made "content-addressable memory" concrete: you recall a whole by presenting a part, and the system completes it by settling. Each stored memory is a basin of attraction; recall is relaxation. Push the state too far between two memories and it can fall into a spurious blend — the network's version of a false or confabulated recollection, which is itself telling.
I keep it here because it quietly dissolves a bad intuition about minds: that remembering must be looking something up in a store, with an address and a librarian. Here there is no store, no librarian, no place where the memory "is" — only a dynamics that, prompted, falls into a familiar shape: no headquarters, no central reader — coherence as what a system settles into, not what a part contains.
Source: Hopfield, PNAS (1982) ↗