In his paper “Pandemonium: A Paradigm for Learning,” presented at the 1958 Mechanisation of Thought Processes symposium and published in the 1959 proceedings, Oliver Selfridge proposed a way for a machine to learn to recognize patterns. He described the system as a crowd of small independent processes he playfully called demons, each shouting with a strength that depended on how strongly it detected its particular feature, with higher-level demons listening to the ones below and a decision demon picking the loudest interpretation.
The architecture had layers. At the bottom, data demons held the raw input. Above them, computational or feature demons looked for specific features and called out when they found them. Cognitive demons combined those features into evidence for whole patterns, and a decision demon chose the answer. Crucially, the system could learn: the weights connecting demons could be adjusted so that useful features counted for more, and Selfridge sketched how the collection of feature detectors might itself be improved over time.
Pandemonium belongs in the prehistory of artificial intelligence because its picture of recognition, many simple parallel units, organized in layers, with adjustable weights and a learning rule, is recognizably the shape of later neural networks. It ran in parallel to Frank Rosenblatt’s perceptron work and helped establish that perception could be modeled as distributed, trainable computation rather than fixed rules. The primary source used here is the archived paper hosted by AITopics.
For a general reader, Pandemonium is an early articulation of an idea now everywhere in AI: that recognizing something is best done not by one clever rule but by many small detectors whose votes are learned and combined.