[Paper Review] Learning and Evaluating Representations for Deep One-class Classification

Kihyuk Sohn, Chunliang Li+3 more|arXiv (Cornell University)|Nov 4, 2020

Anomaly Detection Techniques and Applications77 references92 citations

TL;DR

A two-stage framework learns self-supervised representations from one-class data and then trains one-class classifiers (OC-SVM or KDE) on those representations, with distribution-augmented contrastive learning improving performance on visual one-class benchmarks.

ABSTRACT

We present a two-stage framework for deep one-class classification. We first learn self-supervised representations from one-class data, and then build one-class classifiers on learned representations. The framework not only allows to learn better representations, but also permits building one-class classifiers that are faithful to the target task. We argue that classifiers inspired by the statistical perspective in generative or discriminative models are more effective than existing approaches, such as a normality score from a surrogate classifier. We thoroughly evaluate different self-supervised representation learning algorithms under the proposed framework for one-class classification. Moreover, we present a novel distribution-augmented contrastive learning that extends training distributions via data augmentation to obstruct the uniformity of contrastive representations. In experiments, we demonstrate state-of-the-art performance on visual domain one-class classification benchmarks, including novelty and anomaly detection. Finally, we present visual explanations, confirming that the decision-making process of deep one-class classifiers is intuitive to humans. The code is available at https://github.com/google-research/deep_representation_one_class.

Motivation & Objective

• Motivate and address the challenge of learning high-quality representations for one-class classification.
• Propose a two-stage framework separating representation learning from one-class classifier training.
• Systematically compare self-supervised methods (augmentation prediction, contrastive learning) for one-class tasks.
• Introduce distribution-augmented contrastive learning to improve feature representations for one-class detection.
• Demonstrate state-of-the-art results on visual one-class benchmarks and provide visual explanations.

Proposed method

• Two-stage framework: (1) learn representations from one-class data using self-supervised methods with a projection head to improve downstream utility; (2) train a one-class classifier (OC-SVM or KDE) on the learned representations.
• Analyse limitations of vanilla contrastive learning for one-class tasks (uniformity and class collision) and propose fixes.
• Introduce distribution-augmented contrastive learning that expands training distributions via geometric transformations to reduce uniformity and improve separability.
• Evaluate multiple self-supervised methods (augmentation prediction, rotation prediction, vanilla and distribution-augmented contrastive learning) and compare OC-SVM vs KDE as detectors.
• Provide a gradient-based explanation approach for deep one-class classifiers, using KDE over learned representations to enable visual explanations.
• Apply the approach to datasets including CIFAR-10/100, Fashion-MNIST, Cat-vs-Dog, CelebA, and MVTec AD, with extensive ablations.

Experimental results

Research questions

• RQ1 How can self-supervised representations be learned effectively from one-class data for reliable one-class classification?
• RQ2 What is the impact of different self-supervised tasks (augmentation prediction, rotation, contrastive learning) on one-class detection performance?
• RQ3 Does distribution-augmented contrastive learning improve one-class representation quality over vanilla methods?
• RQ4 How do OC-SVM and KDE compare as detectors when using learned representations for one-class tasks?
• RQ5 Can we provide meaningful gradient-based explanations for deep one-class classifiers?

Key findings

• The distribution-augmented contrastive learning method achieves the highest mean AUC (89.9) across multiple datasets, outperforming vanilla methods.
• Using KDE or OC-SVM on learned representations yields competitive or superior results compared to surrogate classifiers trained on simulated outliers.
• Rotation prediction with an MLP projection head improves AUC, and KDE on RotNet representations can surpass RotNet with a direct rotation classifier.
• Contrastive learning alone underperforms without fixes due to uniformity and class collision; proper adjustments (limited batch size, projection head, distribution augmentation) boost performance.
• The two-stage framework yields state-of-the-art results on CIFAR-10, CIFAR-100, Fashion-MNIST, Cat-vs-Dog, CelebA, and MVTec AD without test-time augmentation or ensembles.