Detecting unusual input to neural networks

TL;DR

This paper presents a method for detecting unusual inputs to neural networks by evaluating their informative content, helping identify when a network might produce unreliable predictions on unfamiliar data.

Contribution

The authors introduce a simple, effective technique to assess input unusualness by comparing informative content, facilitating direct comparison of uncertainty metrics across different scales.

Findings

The proposed method effectively detects out-of-distribution inputs.

It allows direct comparison of uncertainty metrics regardless of scale.

The approach outperforms existing methods in various datasets.

Abstract

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.