Programming Machine Learning

Where Perceptrons Fail

What’s not to love about perceptrons? They’re simple, and they can be assembled into larger structures like machine learning construction bricks. However, that simplicity comes with a distressing limitation: perceptrons work well on some datasets, and fail badly on others. More specifically, perceptrons are a good fit for linearly separable data. Let’s see what “linearly separable” means, and why it matters.

Linearly Separable Data

Take a look at this two-dimensional dataset:

The two classes in the data—green triangles and blue squares—are neatly arranged into distinct clusters. You could even separate them with a line, ...