Chapter 20. Neural Networks
At the heart of neural networks is the unit (also called a node or neuron). A unit takes in one or more inputs, multiplies each input by a parameter (also called a weight), sums the weighted input’s values along with some bias value (typically 0), and then feeds the value into an activation function. This output is then sent forward to the other neurals deeper in the neural network (if they exist).
Feedforward neural networks—also called multilayer perceptron—are the simplest artificial neural network used in any real-world setting. Neural networks can be visualized as a series of connected layers that form a network connecting an observation’s feature values at one end, and the target value (e.g., observation’s class) at the other end. The name feedforward comes from the fact that an observation’s feature values are fed “forward” through the network, with each layer successively transforming the feature values with the goal that the output at the end is the same as the target’s value.
Specifically, feedforward neural networks contain three types of layers of units. At the start of the neural network is an input layer where each unit contains an observation’s value for a single feature. For example, if an observation has 100 features, the input layer has 100 nodes. At the end of the neural network is the output layer, which transforms the output of the hidden layers into values useful for the task at hand. For example, if our goal ...