Multi-layer perceptrons are in a class of machine learning models called neural networks. The term arose among early computer scientists who understood that vast numbers of highly-interconnected neurons comprised mammalian brains. They envisioned the benefits of simulating the operation of neurons in computers. Indeed, we are now seeing the power of neural networks to solve problems recently considered intractable. However, modern computer scientists do not claim that neural networks simulate brains with great fidelity.
In neural networks for machine learning, a neuron refers to a model that composes a linear combination of inputs with a nonlinear function. The logistic regression model is a simple neuron, but other nonlinear activation functions can be used for various applications. The parameters are usually called the weights of the model, because the linear combination is a weighted sum of the inputs.
Neural networks can solve difficult problems with high accuracy when numerous neurons are organized in layers. The muli-layer perceptron is a simple example, having a feature layer and an output layer. Layers between the input and output are often called “hidden layers.” With notable exceptions, each layer generates output from values in the previous layer. Such networks are therefore called “feedforward neural networks.”
The number of layers, the number of neurons in each layer, and the choice of nonlinear functions are all hyperparameters of a neural network and comprise its architecture or topology.
The process of calculating the output of a neural network from a given input is frequently called forward propagation, because calculations flow forward from the inputs, through intermediate layers, to the output.
The process of adjusting the parameters during training is frequently called backward propagation or “backprop,” because calculations can be envisioned to start with differences between the model output and the training output, and then flow backward.
Neural networks with three or more layers are called “deep neural networks.” They were time consuming to study and train before the advent of modern computers and fast linear algebra libraries, because of the large number of calculations involved. However, it is deep neural networks that have spawned resurgent interest in artificial intelligence by their ability to solve computer vision, speech recognition and other types of problems.