In the summer of 1951, while Marvin Minsky was a graduate student, he and Dean Edmonds built one of the first machines to physically implement a neural network. They called it the SNARC, the Stochastic Neural-Analog Reinforcement Calculator. Funded by a small grant from the Office of Naval Research, it used around 300 vacuum tubes together with motors and clutches to realize about 40 artificial “neurons,” each with an adjustable connection strength stored on a mechanical control.
The machine simulated a rat learning to run a maze. A signal representing the rat moved through the network; when it made choices that led toward the goal, the connections that had contributed were strengthened, so that good paths became more likely on later runs - a hardware form of reinforcement learning loosely inspired by Hebbian ideas about synapses. Minsky later recalled the machine’s quirks, including that its random wiring made it robust: if a neuron failed, it barely mattered, and the maze could even be run with more than one simulated rat at a time.
The SNARC sits alongside Shannon’s Theseus and Turing’s “unorganised machines” as an early attempt to build, rather than just describe, a learning machine made of neuron-like parts. Minsky himself would later become famous both as a founder of AI and, with Seymour Papert’s 1969 book Perceptrons, as a sharp critic of the limits of simple neural networks - but his own first machine was a neural net.