Likelihood of generative models has been used traditionally as a score to detect atypical (Out-of-Distribution, OOD) inputs. However, several recent studies have found this approach to be highly unreliable, even with invertible generative models, where computing the likelihood is feasible. In this paper, we present a different framework for generative model--based OOD detection that employs the model in constructing a new representation space, instead of using it directly in computing typicality scores, where it is emphasized that the score function should be interpretable as the similarity between the input and training data in the new space. In practice, with a focus on invertible models, we propose to extract low-dimensional features (statistics) based on the model encoder and complexity of input images, and then use a One-Class SVM to score the data. Contrary to recently proposed OOD detection methods for generative models, our method does not require computing likelihood values. Consequently, it is much faster when using invertible models with iteratively approximated likelihood (e.g. iResNet), while it still has a performance competitive with other related methods

Detecting novelties given unlabeled examples of normal data is a challenging task in machine learning, particularly when the novel and normal categories are semantically close. Large deep models pretrained on massive datasets can provide a rich representation space in which the simple k-nearest neighbor distance works as a novelty measure. However, as we show in this paper, the basic k-NN method might be insufficient in this context due to ignoring the 'local geometry' of the distribution over representations as well as the impact of irrelevant 'background features'. To address this, we propose a fully unsupervised novelty detection approach that integrates the flexibility of k-NN with a locally adapted scaling of dimensions based on the 'neighbors of nearest neighbor' and computing a 'likelihood ratio' in pretrained (self-supervised) representation spaces. Our experiments with image data show the advantage of this method when off-the-shelf vision transformers (e.g., pretrained by DINO) are used as the feature extractor without any fine-tuning.

Contrastive Predictive Coding (CPC) is a well-established self-supervised learning method that naturally fits time series data. This method has been recently leveraged to detect anomalous inputs, viewed as the task of classifying positive pairs of context-feature representations versus negative ones in order to employ classifier uncertainty measures. In this paper, by taking a different perspective, we propose a CPC-based Out-Of-Distribution (OOD) detection method for time series data that does not require any negative samples at test time and is theoretically related to a probabilistic type of uncertainty estimation in the latent representation space. Our method extends the standard CPC by using a radial (distance-based) score function both in the training loss and as the OOD measure, in addition to quantizing the context (replacing it by cluster prototypes) during inference. The proposed method is applied to detecting OOD human activities with smartphone sensors data and shows promising performance on two primary datasets without using activity labels in training.

A deep neural classifier trained on an upstream task can be leveraged to boost the performance of another classifier in a related downstream task through the representations learned in hidden layers. However, presence of shortcuts (easy-to-learn features) in the upstream task can considerably impair the versatility of intermediate representations and, in turn, the downstream performance. In this paper, we propose a method to improve the representations learned by deep neural image classifiers in spite of a shortcut in upstream data. In our method, the upstream classification objective is augmented with a type of adversarial training where an auxiliary network, so called lens, fools the classifier by exploiting the shortcut in reconstructing images. Empirical comparisons in self-supervised and transfer learning problems with three shortcut-biased datasets suggest the advantages of our method in terms of downstream performance and/or training time.