The perceptron: A probabilistic model for information storage and organization in the brain

Inside the paper

The problem

Rosenblatt asks how a system might retain experience and use it to recognize new sensory patterns. He examines a model in which memory lies in associations rather than a stored copy of each stimulus.

The idea

The perceptron connects sensory inputs to association units and then to response units. Reinforcement changes the influence of active associations, allowing experience to affect later responses.

How it works

• A stimulus activates a pattern on a sensory surface, which sends signals through connections to association units.
• Association units respond to combinations of sensory activity and transmit their activity toward possible responses.
• Reinforcement changes the values associated with active connections, altering the balance of support for later responses.
• The analysis compares random stimulus environments with differentiated environments in which members of a class share structure.

What it showed

The theoretical analysis finds that performance depends on the organization of stimuli and the system’s parameters. In the differentiated setting, the model can assign previously unseen patterns to the appropriate class above chance under the paper’s stated conditions.

Read it for

Begin with the model description and Figures 1 and 2 on paper pages 389–390, then read the conclusions on pages 404–405. The probability derivations between them can be skimmed first and revisited for the assumptions behind the claims.

Written in our own words from the paper itself. A draft the editors have not yet reviewed.