A perceptron is an early artificial neuron model and linear binary classifier, introduced by psychologist Frank Rosenblatt in 1958, widely regarded as one of the foundational building blocks of the Neural network and connectionist tradition in Artificial intelligence. It computes a weighted sum of its inputs, adds a bias term, and passes the result through a step function to output one of two classes.
Mechanism
Rosenblatt's perceptron takes a vector of numerical inputs, multiplies each by an adjustable weight, sums the results, and compares the total against a threshold to produce a binary output. Weights are adjusted through a simple supervised learning rule, the perceptron learning algorithm, which nudges weights toward reducing classification error on labeled training examples. The perceptron convergence theorem guarantees that this procedure will find a separating solution in a finite number of steps, but only if the training data is linearly separable, meaning a single straight line, or hyperplane in higher dimensions, can divide the two classes.
The 1958 unveiling
Rosenblatt built the perceptron at the Cornell Aeronautical Laboratory with funding from the U.S. Office of Naval Research, and demonstrated an early hardware implementation, the Mark I Perceptron, using a grid of photocells as visual input. Press coverage at the time, including a 1958 New York Times article, described claims that the machine would eventually be able to "walk, talk, see, write, reproduce itself and be conscious of its existence," an early instance of AI hype that later contributed to disappointment when the technology's limitations became apparent.
The Minsky-Papert critique
In 1969, Marvin Minsky and Seymour Papert published "Perceptrons," a rigorous mathematical analysis demonstrating that single-layer perceptrons could not solve problems that are not linearly separable, most famously the XOR logical function. Although the book acknowledged that multi-layer networks could in principle overcome this limitation, it was widely interpreted, fairly or not, as a broader case against neural network research, and is often credited with contributing to a sharp decline in funding and interest that lasted into the AI winter of the 1970s and beyond.
Legacy and multilayer perceptrons
Interest in neuron-inspired models revived in the 1980s with the popularization of Backpropagation by David Rumelhart and colleagues, an algorithm that made it practical to train multi-layer perceptrons, networks with one or more hidden layers between input and output, which can represent non-linear decision boundaries and overcome the exact limitation Minsky and Papert identified. Multi-layer perceptron architectures remain a basic building block of modern Deep learning systems, and Rosenblatt's original single-layer model is now taught primarily as a historical and pedagogical starting point for understanding Gradient descent-based learning and the broader development of Machine learning.