Detecting unusual input to neural networks
This work addresses the issue of unreliable predictions in neural networks for users dealing with out-of-distribution data, but it appears incremental as it builds on existing uncertainty evaluation techniques.
The paper tackles the problem of detecting unusual inputs to neural networks that differ from training data, which can cause flawed predictions, by introducing a method to judge input unusualness based on informative content compared to learned parameters, and it compares this approach to existing uncertainty evaluation methods across various datasets and scenarios.
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 input by evaluating its informative content compared to the learned parameters. This technique can be used to judge whether a network is suitable for processing a certain input and to raise a red flag that unexpected behavior might lie ahead. We compare our approach to various methods for uncertainty evaluation from the literature for various datasets and scenarios. Specifically, we introduce a simple, effective method that allows to directly compare the output of such metrics for single input points even if these metrics live on different scales.