Role of perceptron in neural network

A perceptron (Rosenblatt, 1958) is the simplest neural network: a single artificial neuron that computes a weighted sum of its inputs and applies a threshold (step) activation to give a binary output.

y = 1 if Σ wᵢxᵢ + b > 0, otherwise 0 (or −1).

 x1 --w1-->\
 x2 --w2--->( Σ + b )--> step --> y
 xn --wn-->/

Role of the perceptron in neural networks

1. Basic building block: every unit of a multilayer network is a generalisation of the perceptron (with a smooth activation); networks of perceptrons form multilayer perceptrons (MLP).
2. Binary classifier: it separates two classes with a linear decision boundary (a line or hyperplane), so it solves linearly separable problems such as AND and OR.
3. Learning model: the perceptron learning rule w ← w + η(t − y)x was one of the first supervised learning algorithms. It converges if the classes are linearly separable.
4. Logic gates: a single perceptron can implement AND, OR and NOT.
5. Historic role: its limitation (it cannot learn XOR) led to multilayer networks and back-propagation.

Limitation: a single perceptron cannot solve non-linearly separable problems; hidden layers are needed.