Summing Up

This linear-separability issue was one of the first to be seriously studied in early machine learning. Nowadays, such a problem is extremely trivial for any major method in the field.

Ultimately, perceptrons are just too weak, for all their nice properties, to solve very ambitious tasks. However, this shouldn’t be taken as a deathblow to the method, or as a statement against the need to study the perceptron. There is a more general class of classifier methods, called the artificial neuron, or AN, of which perceptrons are the first described, simplest, and in many ways weakest. ANs were designed to overcome the limitations of the perceptron. Individually, ANs aren’t much stronger than the perceptron, with the major exception of the SVM. However, all our cutting edge methods, especially what we now call deep learning, are ultimately a major extension and abstraction of the artificial neuron class of methods. Artificial neurons should be seen as building blocks, conceptually and almost literally, of all our current cutting edge methods, and understanding the perceptron is the best first step towards a strong background in machine learning as it is used today.