Overview

The perceptron (and other linear algorithms such as SVM) is designed to recognize linear patterns. When the patterns are not linear, the percetron cannot be used to recognize the pattern without using some technique to linearize the pattern first.

Demo

The following is a graph of a sample non linear pattern. The points divided into two categories. One category is a circular shape, and the second category consists of points outside that circle. As can be seen, no single line can be drawn to separate the two sets of points.

Linearizing a Non Linear Pattern

There are different techniques for turning a non-linear pattern into a linear one.

• Projecting to a Higher Dimensional Space
• Kernels -

Non-Linear Algorithms

While linearly separable learning algorithms can be made effective, there are algorithms that are not limited to learning linear patterns.

• Neural Networks