Artificial neural networks
Networks of simple threshold units can compute and learn from examples.
Mathematics and information · Machine learning · No. 125
- Electrical and chemical age
- Science
A 1943 model that only added up inputs and fired led to a 1958 machine that taught itself to sort cards.
Chart: Illustration: Artificial neural networks
Where it comes from
Chart: Map of where it comes from
Warren McCulloch and Walter Pitts published their logical model of the neuron in 1943. Frank Rosenblatt turned it into a learning machine, the perceptron, which the US Navy unveiled in July 1958.
Three things to know
- McCulloch and Pitts used the all-or-none firing of nerve cells to treat neural events with propositional logic.
- In July 1958 the US Office of Naval Research unveiled Rosenblatt's perceptron, which taught itself to tell cards marked on the left from cards marked on the right.
- A single layer of such units cannot solve the XOR problem, and the limit is blamed for the AI winter of the 1970s.
The journey
- 1 Nerve cellFires all or nothing when its inputs add up enough
- 2 ModelMcCulloch and Pitts treat neurons with logic, 1943
- 3 LearningRosenblatt's perceptron adjusts its connections, 1958
- 4 SetbackMinsky and Papert show the limits of one layer, 1969
- 5 RevivalMany-layer networks trained by backpropagation, 1986
- 6 TodayVision, speech, robots and language models