Abstract:Anomaly detection aims at identifying deviant samples from the normal data distribution. Much progress has been made in recent years for anomaly detection with self-supervised representation learning. However, most existing approaches assume the training set contains either only clean normal samples or some labeled abnormal samples. With a contaminated unlabeled training set, the performance is degraded with unclear discrimination between normal and abnormal samples. To address the above challenge, in this paper, we propose a novel unsupervised representation learning framework that takes advantage of the extra information provided by the anomalies in the unlabeled contaminated data for anomaly detection. Specifically, anomaly discrimination learning with pseudo normality score generation and multi-instance contrastive learning is proposed for distinguishing abnormal samples from the unlabeled set. Meanwhile, we combine the discrimination learning with mean-shifted contrastive learning and distribution-shifting transformation classification into a multi-task representation learning framework to improve the stability and robustness of the training. Our proposed method achieves state-of-the-art performance on CIFAR-10, ImageNet-10, MNIST and F-MNIST datasets. Extensive ablation studies further demonstrate the effectiveness of different components of the proposed framework.

Multi-Instance Learning with Distribution Change

Multi-Instance Learning with Key Instance Shift

Multiple-Instance Learning from Pairwise Comparison Bags

Learning Discrimination from Contaminated Data: Multi-Instance Learning for Unsupervised Anomaly Detection

Multi-Instance Learning: A Survey

Neural Networks for Multi-Instance Learning

Multi-instance learning by treating instances as non-I.I.D. samples

Multi-Instance Causal Representation Learning for Instance Label Prediction and Out-of-Distribution Generalization

Multi-Instance Learning from Supervised View

Ensembles Of Multi-Instance Learners

Solving the Multi-Instance Problem with Neural Networks

Multi-instance clustering with applications to multi-instance prediction

A Manifold Learning-Based Multi-Instance Regression Algorithm

A New Multiple Instance Algorithm Using Structural Information.

A New Multi-Instance Learning Scheme For Scene Categorization Using Information Bottleneck Theory

Multi-instance Kernel Learning with Concept Weights of Instance Space

Multiple instance learning from similarity-confidence bags

Multiple instance learning with bag dissimilarities

Adaptive P-Posterior Mixture-Model Kernels for Multiple Instance Learning

Multi-label learning by instance differentiation

Multi-Instance Multi-Label Learning with Application to Scene Classification