perceptron: a standalone construct

Frank Rosenblatt's 1958 paper The perceptron: A probabilistic model for information storage and organization in the brain introduced the perceptron, the first trainable artificial neuron: a unit that takes several numerical inputs, multiplies each by an adjustable weight, sums them, and fires a yes-or-no output when the total passes a threshold. Its defining feature is the accompanying learning rule, a simple procedure that nudges the weights up or down each time the unit misclassifies a labelled example, so the machine improves from experience rather than from hand-coded instructions; Rosenblatt's 1962 Principles of neurodynamics set out the fuller theory. Because a single perceptron draws only a straight-line (linear) boundary between two categories, Marvin Minsky and Seymour Papert's 1969 book Perceptrons showed that one such unit cannot compute problems that are not linearly separable, the exclusive-or function being the standard example. The construct is a landmark as the earliest supervised, weight-adjusting classifier, the direct ancestor of the multi-layer neural networks and deep-learning models that stack many such units together.[1][2][3]

Unit

A single computational unit defined by its input weights, a threshold, and a binary output

Recovery means

Not applicable

Measured by

Specified by its weight vector and threshold; performance assessed by classification accuracy on labelled data, with convergence of the learning rule guaranteed only when the classes are linearly separable