JOVANA
Library Glossary Getting Started Three Levels Fields How it works Mission
Join the mission
Back to the library
Artificial Intelligence 1982

Neural Networks and Physical Systems with Emergent Collective Computational Abilities

John J. Hopfield

Wire neurons symmetrically and memories become valleys; a fragment rolls downhill to the whole.

Choose your version
In depth · the introduction

Show this network a smudged, half-erased letter and it hands you back the whole letter — clean. It remembers the way memory should: by content, not by address.

The idea, unpacked

Ordinary computer memory is like a wall of numbered lockers: to get something out, you need its address. Human memory is nothing like that — a few bars of a tune, a glimpse of a face, and the whole thing comes back. In 1982 the physicist John Hopfield built a small network of neurons that worked the second way.

Picture a hilly landscape with a few deep valleys, one for each thing the network has learned. Hand it a clue — a corrupted or partial pattern — and you've dropped a ball somewhere on a hillside. The network's rule is simply that the ball always rolls downhill, and it comes to rest at the bottom of the nearest valley: the complete, remembered pattern. Memory, in this picture, is just rolling downhill to the nearest stable state.

Where it came from

Hopfield was not a neuroscientist by training but a physicist, already celebrated for work on how light moves through crystals and for "kinetic proofreading" in biology. He came to the brain through the physics of spin glasses — disordered magnets whose atoms, like tiny compass needles, settle into low-energy arrangements. He saw that a network of neurons with symmetric connections is the same kind of system, and that its "settling" could be the reading-out of a memory.

The little paper landed at the right moment. Neural networks had been in a long winter since the 1969 critique of the perceptron, and Hopfield's clean physical picture — memories as energy valleys — helped pull the field back to life in the 1980s. Forty-two years later, in 2024, that revival was recognized with the Nobel Prize in Physics, shared by Hopfield and Geoffrey Hinton, whose Boltzmann machine grew directly out of this model.

Why it mattered

Hopfield showed that memory and computation need not be programmed in — they can emerge, on their own, from a crowd of simple parts all obeying one local rule. That single idea — let a physical system seek its lowest energy and read the answer off where it stops — runs underneath a surprising amount of modern AI, from the energy-based models of the 1980s to the attention mechanism inside today's language models. And because no single neuron is essential, the memory degrades gently when parts fail, just like ours.

Like a marble in an egg carton

Imagine rolling a marble across an egg carton. Wherever you let it go, it rolls into the nearest cup and stops. Each cup is a memory the network has stored; the spot where you drop the marble is your clue. A clue near the right cup lands in it and recovers the whole memory; a clue thrown too wildly can fall into the wrong cup instead. Try it below: pick a letter, smudge it with the noise slider, then let the network roll it back into shape.

A 5×5 grid of neurons holding one of three stored letters (T, C, H). The noise slider flips a fixed set of pixels to corrupt the cue; the recall slider replays the asynchronous threshold dynamics sweep by sweep. Mismatched pixels are ringed in red and vanish as the network settles; with light noise it reconstructs the exact stored letter, while heavy noise can push it into a neighbouring or spurious valley.

What came before and after

The network's all-or-none neurons descend from the 1943 McCulloch–Pitts model, and its learning rule is Donald Hebb's 1949 "cells that fire together, wire together." It sits beside the other founders of machine learning in this Library — the perceptron (Rosenblatt 1958), and downstream, AlexNet (2012) and the Transformer (2017). The thread is direct: Hopfield's energy idea became the Boltzmann machine, which helped revive deep learning, and the modern version of his recall rule turns out to be the very attention operation the Transformer runs on.

The original document
Original source text
J. J. Hopfield · Divisions of Chemistry and Biology, California Institute of Technology, and Bell Laboratories · PNAS 79(8), 2554–2558 · April 1982 (Biophysics; contributed January 15, 1982)
The opening claim
Computational properties of use to biological organisms or to the construction of computers can emerge as collective properties of systems having a large number of simple equivalent components (or neurons).
From this premise Hopfield builds a model of memory not as a set of addressed locations but as the collective behaviour of many identical two-state neurons — its computational power an emergent property of the whole.
[ … ]
The model
He defines N two-state neurons, fully interconnected by symmetric weights Tij = Tji (with Tii = 0) set by a Hebbian outer-product rule over the stored states. The state evolves as each neuron asynchronously compares its weighted input against a threshold and switches on or off accordingly.
The algorithm for the time evolution of the state of the system is based on asynchronous parallel processing.
The paper's pivotal step is to exhibit an energy function E = −½ Σ Tij Vi Vj that the symmetric dynamics can only decrease. The stored patterns are arranged to be its local minima, so the flow toward minimum energy reconstructs a complete memory from a partial cue — content-addressable memory. (The energy function, the convergence argument, the capacity simulations, and the integrated-circuit realization are given in full at the source.)
[ … ]
What emerges
Additional emergent collective properties include some capacity for generalization, familiarity recognition, categorization, error correction, and time sequence retention.
Hopfield reports that the memory is robust: it is only weakly sensitive to the modeling details or to the failure of individual neurons, and his simulations indicate roughly 0.15N patterns can be stored before recall degrades — a number a later spin-glass analysis sharpened to about 0.138N.
California Institute of Technology & Bell Laboratories · 1982