Characterizing out-of-distribution generalization of neural networks: application to the disordered Su–Schrieffer–Heeger model

Abstract Machine learning (ML) is a promising tool for the detection of phases of matter. However, ML models are also known for their black-box construction, which hinders understanding of what they learn from the data and makes their application to novel data risky. Moreover, the central challenge of ML is to ensure its good generalization abilities, i.e. good performance on data outside the training set. Here, we show how the informed use of an interpretability method called class activation mapping, and the analysis of the latent representation of the data with the principal component analysis can increase trust in predictions of a neural network (NN) trained to classify quantum phases. In particular, we show that we can ensure better out-of-distribution (OOD) generalization in the complex classification problem by choosing such an NN that, in the simplified version of the problem, learns a known characteristic of the phase. We also discuss the characteristics of the data representation learned by a network that are predictors of its good OOD generalization. We show this on an example of the topological Su–Schrieffer–Heeger model with and without disorder, which turned out to be surprisingly challenging for NNs trained in a supervised way. This work is an example of how the systematic use of interpretability methods can improve the performance of NNs in scientific problems.