SNARC, the first neural network machine

In 1951 two Harvard graduate students, Marvin Minsky and Dean Edmonds, spent a grant of a few thousand dollars on 300 vacuum tubes, a bank of motors, and the automatic pilot from a B-24 bomber. Out of it they built the Stochastic Neural Analog Reinforcement Calculator, forty artificial neurons whose connection strengths were set by potentiometers, turned by motors, driven by a clutch that engaged whenever the machine was rewarded.

The task was a rat in a maze. The machine chose moves at random; when a run reached the goal, the clutch tightened the connections that had just been active, so that the same choices became more likely next time. That is Hebb's rule with a reward signal added, and it is the seed of what would later be called reinforcement learning. Minsky reported that the machine learned, and also that it kept learning after a couple of neurons burned out, which he took as a sign that distributed representations were robust.

Minsky wrote his 1954 doctoral thesis on neural-analogue reinforcement systems, then spent the next fifteen years arguing that the symbolic route was more promising. The man who built the first network machine became the author of the book that put networks in the cold.