R Deep Learning Essentials - Second Edition

Generating predictions from a neural network

For any given observation, there can be a probability of membership in any of a number of classes (for example, an observation may have a 40% chance of being a 5, a 20% chance of being a 6, and so on). To evaluate the performance of the model, some choices have to be made about how to go from the probability of class membership to a discrete classification. In this section, we will explore a few of these options in more detail.

As long as there are no perfect ties, the simplest method is to classify observations based on the highest predicted probability. Another approach, which the RSNNS package calls the winner takes all (WTA) method, chooses the class with the highest probability, provided the ...