Multilayer Perceptrons

A multilayer perceptron (MLP) is a feedforward artificial neural network model that maps sets of input data onto a set of appropriate outputs.

Source: https://en.wikipedia.org/wiki/Multilayer_perceptron

Notation

You can write the topology (architecture) of a MLP by writing down the number of neurons per layer separated by ‘:’. Three examples are:

• 160:10 - A MLP with 160 input neurons (160 features) and 10 output neurons. It has no hidden layer(s).
• 160:1337:456 - A MLP with 160 input neurons, one hidden layer with 1337 neurons and 456 output neurons (classes)
• 4:20:21:4 - A MLP with 4 input neurons, 20 neurons in the first hidden layer and 21 in the second hidden layer. It has 4 output neurons.

If you want to show the activation function like this:

(160,sigmoid):(500,sigmoid):(500,tanh):(369,softmax)

If it is not noted otherwise, the activation function is always sigmoid and in the last layer softmax.

Classification

Classification tasks can be tackled with MLPs by using one input neuron per feature and one output neuron per class. For example, if you have a 28×28 pixel image which is a digit (0, 1, 2, 3, 4, 5, 6, 7, 8, 9), then you would make a MLP with 28 · 28 = 784 input neurons and 10 output neurons.