Detecting unusual input to neural networks

Jörg Martin, Clemens Elster

Evaluating a neural network on an input that differs markedly from the training data might cause erratic and flawed predictions. We study a method that judges the unusualness of an…