This is a huge topic in machine learning, so we can't cover everything in this chapter. If you've never seen a neural network before, they look like a giant spider web. The vertices of these spider webs are called neurons, or units, and they are based on an old-school linear classifier known as a perceptron. The idea is that your vector comes in, computes a dot product with a corresponding weight vector of parameters, and then gets a bias value added to it. Then, we transform it via an activation function. A perceptron, in general, can be canonically the same as logistic regression if you're using a sigmoid transformation.
When you string a whole bunch of these together, what you get is the massive web of perceptrons feeding perceptrons: this is called a multi layer perceptron, but it's also known as a neural network. As each of these perceptrons feeds the next layer, the neurons end up learning a series of nonlinear transformations in the input space,...