Methods for deep learning model failure detection and model adaption: A survey

Abstract

In real-world applications, deep learning models may fail to predict due to service switch, system upgrade, or other environmental changes. One main reason is that the model lacks generalization ability when data distribution changes. To detect model failures in advance, a direct and effective method is to monitor the data distribution in real time. This paper provides a taxonomy of data distribution shift detection methods, which is an important issue in model failure perception, and also gives a framework on model adaption and generalization under distribution shift scenario.