Learning Data Mining with Python - Second Edition

Back-propagation

Training a neural network is specifically focused on the following things.

• The first is the size and shape of the network - how many layers, what sized layers and what error functions they use. While types of neural networks exists that can alter their size and shape, the most common type, a feed-forward neural network, rarely has this capability. Instead, its size is fixed at initialization time, which in this chapter is 400 neurons in the first layer, 100 in the hidden layer and 26 in the final layer. Training for the shape is usually the job of a meta-algorithm that trains a set of neural networks and determines which is the most effective, outside of training the networks themselves.
• The second part of training a neural ...