Until now, you have seen how a perceptron performs binary classification. But if that were the only task a perceptron (or a collection of them) could do, we wouldn’t have cared much about them. It turns out that they can do much more complex things, such as multiclass classification. Let’s see how a set of perceptrons can perform more complex tasks.
We see how a network of perceptrons can act as a universal function approximator. We have seen how a single layer of perceptron in combination with an AND gate leads to an enclosure in a polygon, and multiple such AND outputs using an OR gate lead to an enclosure in multiple polygons. In the most extreme case, this can be extended to finding a polygon for every single data point.
Let’s now test what you have understood in the previous few lectures.
In the following figure, there are two classes shown in two colours.
You know that a single perceptron is a binary classifier and that it can be defined in any of the following shown ways. You have the freedom to decide which one to use to answer the following questions.
That brings us to the end of understanding how networks of simple perceptrons can act as universal function approximators.
Report an error