Definition
Rosenblatt Perceptron
The Rosenblatt perceptron, introduced by Frank Rosenblatt in 1958, is a layered connectionist machine and the first artificial neural network capable of learning from examples.
Its architecture consists of three stages:
- S-cells (sensory): receive input from a retina-like sensor array.
- A-cells (association): groups of S-cells combine onto cells in an association area A1; groups of A1 cells further combine into A2. These are the hidden layers.
- R-cells (response): signals from A2 cells converge onto output cells that produce the final classification.
All connections may be excitatory or inhibitory. Feedback connections run from R-cells back to A-cells, enforcing mutually exclusive outputs — when one response fires, it suppresses competing responses.
Learning
Unlike the McCulloch–Pitts model, the Rosenblatt perceptron could learn. Weights between A-cells and R-cells were adjusted using a variant of Hebbian learning: when the output was wrong, active connections were strengthened or weakened to reduce the error. This is the ancestor of the modern perceptron algorithm.
The perceptron was implemented in hardware as the Mark I Perceptron — a physical machine with a 20 × 20 pixel camera and motor-driven potentiometers as adjustable weights.