Perceptron: the main component of neural networks

The history of the perceptron is interesting. It was invented all the way back in 1957, and seemed like a promising way to learn a function. In fact, it was so promising, that it was actually intended to be implemented in hardware as a physical array of electronic perceptrons. But a seminal paper published in the 60s by Marvin Minsky proved that the perceptron was fatally flawed: it could only learn linearly separable patterns, meaning it couldn't even learn the XOR function.

So essentially neural network research completely came to a halt until a decade later when the idea of a multi-layer perception, which can learn non-linearly separable functions, (and is now the foundation of all modern ANN architectures) became more widely appreciated. (Actually, the original Minsky paper had mentioned that a multi-layer perceptron would overcome the limitations of the single-layer model, but the blow to the original perceptron was apparently so devastating that nobody cared.)