Neural network learning device, method, and program

This invention describes a method for training neural networks, especially when there isn't much training data. It works by extracting key features from the available data and then creating slightly altered, "adversarial" versions of those features. The network then learns by comparing its recognition results using both the original and adversarial data to a desired outcome. The claims narrow this to generating the adversarial features using a specific mathematical constraint derived from the original training data.