A new method for robust and reliable classifications with deep neural networks

Deep neural networks (DNNs) have been widely used for classification tasks in various scientific and biomedical applications, including the prediction of crystal structure properties and disease classification. However, a common drawback of DNNs is their opaque nature, making it difficult to interpret the basis for their classifications. Therefore, it is crucial that predictions are not only accurate but also accompanied by reliable uncertainty estimates. Additionally, these predictions need to be robust since input data can often be noisy, and it is important that minimal changes in the input do not result in drastically different classifications. In this talk, we introduce a novel method that significantly enhances the reliability and robustness of DNNs. These improvements are achieved through a regularizing loss function added to the penultimate layer of the neural network. This constraint forces all latent representations to converge at the vertices of a hypercube, inducing a “binary encoding” of the representations in the latent space. We demonstrate the effectiveness of our method across multiple relevant benchmarking applications.