Recent advances in out-of-distribution (OOD) detection on image data show that pre-trained neural network classifiers can separate in-distribution (ID) from OOD data well, leveraging the class-discriminative ability of the model itself. Methods have been proposed that either use logit information directly or that process the model's penultimate layer activations. With "WeiPer", we introduce perturbations of the class projections in the final fully connected layer which creates a richer representation of the input. We show that this simple trick can improve the OOD detection performance of a variety of methods and additionally propose a distance-based method that leverages the properties of the augmented WeiPer space. We achieve state-of-the-art OOD detection results across multiple benchmarks of the OpenOOD framework, especially pronounced in difficult settings in which OOD samples are positioned close to the training set distribution. We support our findings with theoretical motivations and empirical observations, and run extensive ablations to provide insights into why WeiPer works. Our code is available at: https://github.com/mgranz/weiper.

With the advent of deep learning over the last decade, a considerable amount of effort has gone into better understanding and enhancing Stochastic Gradient Descent so as to improve the performance and stability of artificial neural network training. Active research fields in this area include exploiting second order information of the loss landscape and improving the understanding of chaotic dynamics in optimization. This paper exploits the theoretical connection between the curvature of the loss landscape and chaotic dynamics in neural network training to propose a modified SGD ensuring non-chaotic training dynamics to study the importance thereof in NN training. Building on this, we present empirical evidence suggesting that the negative eigenspectrum - and thus directions of local chaos - cannot be removed from SGD without hurting training performance. Extending our empirical analysis to long-term chaos dynamics, we challenge the widespread understanding of convergence against a confined region in parameter space. Our results show that although chaotic network behavior is mostly confined to the initial training phase, models perturbed upon initialization do diverge at a slow pace even after reaching top training performance, and that their divergence can be modelled through a composition of a random walk and a linear divergence. The tools and insights developed as part of our work contribute to improving the understanding of neural network training dynamics and provide a basis for future improvements of optimization methods.

Attribution methods aim to explain a neural network’s prediction by highlighting the most relevant image areas. A popular approach is to backpropagate (BP) a custom relevance score using modified rules, rather than the gradient. We analyze an extensive set of modified BP methods: Deep Taylor Decomposition, Layer-wise Relevance Propagation (LRP), Excitation BP, PatternAttribution, DeepLIFT, Deconv, RectGrad, and Guided BP. We find empirically that the explanations of all mentioned methods, except for DeepLIFT, are independent of the parameters of later layers. We provide theoretical insights for this surprising behavior and also analyze why DeepLIFT does not suffer from this limitation. Empirically, we measure how information of later layers is ignored by using our new metric, cosine similarity convergence (CSC). The paper provides a framework to assess the faithfulness of new and existing modified BP methods theoretically and empirically.