Deep approaches to anomaly detection have recently shown promising results over shallow methods on large and complex datasets. Typically is treated as an unsupervised learning problem. In practice however, one may have---in addition a set of unlabeled samples---access small pool labeled samples, e.g. subset verified by some domain expert being normal or anomalous. Semi-supervised aim utilize such but most proposed are limited merely including samples. Only few take advantage anomalies, with...

Deep one-class classification variants for anomaly detection learn a mapping that concentrates nominal samples in feature space causing anomalies to be mapped away. Because this transformation is highly non-linear, finding interpretations poses significant challenge. In paper we present an explainable deep method, Fully Convolutional Data Description (FCDD), where the are themselves also explanation heatmap. FCDD yields competitive performance and provides reasonable explanations on common...

Activity coefficients, which are a measure of the nonideality liquid mixtures, key property in chemical engineering with relevance to modeling and phase equilibria as well transport processes. Although experimental data on thousands binary mixtures available, prediction methods needed calculate activity coefficients many relevant that have not been explored date. In this report, we propose probabilistic matrix factorization model for predicting arbitrary mixtures. no physical descriptors...

There exist few text-specific methods for unsupervised anomaly detection, and those that do exist, none utilize pre-trained models distributed vector representations of words. In this paper we introduce a new detection method—Context Vector Data Description (CVDD)—which builds upon word embedding to learn multiple sentence capture semantic contexts via the self-attention mechanism. Modeling enables us perform contextual sentences phrases with respect themes concepts present in an unlabeled...

Though anomaly detection (AD) can be viewed as a classification problem (nominal vs. anomalous) it is usually treated in an unsupervised manner since one typically does not have access to, or infeasible to utilize, dataset that sufficiently characterizes what means "anomalous." In this paper we present results demonstrating intuition surprisingly seems extend deep AD on images. For recent benchmark ImageNet, classifiers trained discern between normal samples and just few (64) random natural...

Due to the intractability of characterizing everything that looks unlike normal data, anomaly detection (AD) is traditionally treated as an unsupervised problem utilizing only samples. However, it has recently been found image AD can be drastically improved through utilization huge corpora random images represent anomalousness; a technique which known Outlier Exposure. In this paper we show specialized learning methods seem unnecessary for state-of-the-art performance, and furthermore one...

A central goal in the cognitive sciences is development of numerical models for mental representations object concepts. This paper introduces Variational Interpretable Concept Embeddings (VICE), an approximate Bayesian method embedding concepts a vector space using data collected from humans triplet odd-one-out task. VICE uses variational inference to obtain sparse, non-negative with uncertainty estimates values. These are used automatically select dimensions that best explain data. We...

Today's computer vision models achieve human or near-human level performance across a wide variety of tasks. However, their architectures, data, and learning algorithms differ in numerous ways from those that give rise to vision. In this paper, we investigate the factors affect alignment between representations learned by neural networks mental inferred behavioral responses. We find model scale architecture have essentially no effect on with responses, whereas training dataset objective...

Recent work has shown that finite mixture models with m components are identifiable, while making no assumptions on the components, so long as one access to groups of samples size 2m -1 which known come from same component.In this we generalize result and show that, if every subset k a model linearly independent, then is identifiable only (2m -1)/(k -1) per group.We further value cannot be improved.We prove an analogous for stronger form identifiability "determinedness" along corresponding...

When estimating finite mixture models, it is common to make assumptions on the components, such as parametric assumptions. In this work, we no distributional components and instead assume that observations from model are grouped, in same group known be drawn component. We precisely characterize number of $n$ per needed for identifiable, a function $m$ components. addition our assumption-free analysis, also study settings where either linearly independent or jointly irreducible. Furthermore,...

Deep neural networks have reached human-level performance on many computer vision tasks. However, the objectives used to train these enforce only that similar images are embedded at locations in representation space, and do not directly constrain global structure of resulting space. Here, we explore impact supervising this by linearly aligning it with human similarity judgments. We find a naive approach leads large changes local representational harm downstream performance. Thus, propose...

Recent research has established sufficient conditions for finite mixture models to be identifiable from grouped observations. These allow the components nonparametric and have substantial (or even total) overlap. This work proposes an algorithm that consistently estimates any model Our analysis leverages oracle inequality weighted kernel density estimators of distribution on groups, together with a general result showing consistent estimation groups implies components. A practical...

Deep anomaly detection is a difficult task since, in high dimensions, it hard to completely characterize notion of "differentness" when given only examples normality. In this paper we propose novel approach deep based on augmenting large pretrained networks with residual corrections that adjusts them the detection. Our method gives rise highly parameter-efficient learning mechanism, enhances disentanglement representations model, and outperforms all existing methods including other baselines...

Regularizing the input gradient has shown to be effective in promoting robustness of neural networks. The regularization input's Hessian is therefore a natural next step. A key challenge here computational complexity. Computing inputs computationally infeasible. In this paper we propose an efficient algorithm train deep networks with operator-norm regularization. We analyze approach theoretically and prove that operator norm relates ability network withstand adversarial attack. give...