Self-supervise, Refine, Repeat: Improving Unsupervised Anomaly Detection

Authors: Jinsung Yoon, Kihyuk Sohn, Chun-Liang Li, Sercan O Arik, Chen-Yu Lee, Tomas Pfister

TMLR 2022 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We conduct extensive experiments across various datasets from different domains, including semantic AD (CIFAR-10 (Krizhevsky & Hinton, 2009), Dog-vs-Cat (Elson et al., 2007)), real-world manufacturing visual AD use case (MVTec (Bergmann et al., 2019)), and real-world tabular AD benchmarks (e.g., detecting medical or network anomalies). We evaluate models at different anomaly ratios of unlabeled training data and show that SRR significantly boosts performance.
Researcher Affiliation Industry EMAIL Google Cloud AI
Pseudocode Yes Algorithm 1 SRR: Self-supervise, Refine, Repeat. Input: Train data D = {xi}N i=1, Ensemble count (K), threshold (γ) Output: Refined data ( ˆD), trained OCC (f), feature extractor (g)
Open Source Code No The paper does not contain an explicit statement about the release of source code for the methodology described, nor does it provide a direct link to a code repository.
Open Datasets Yes We conduct extensive experiments across various datasets from different domains, including semantic AD (CIFAR-10 (Krizhevsky & Hinton, 2009), Dog-vs-Cat (Elson et al., 2007)), real-world manufacturing visual AD use case (MVTec (Bergmann et al., 2019)), and real-world tabular AD benchmarks (e.g., detecting medical or network anomalies). Following (Zong et al., 2018; Bergman & Hoshen, 2019), we test the performance of SRR on a variety of real-world tabular AD datasets, including network (KDDCup) and medical (Thyroid, Arrhythmia) AD from the UCI repository (Asuncion & Newman, 2007).
Dataset Splits Yes To construct the data splits, we utilize 50% of normal samples for training. In addition, we hold out some anomaly samples (amounting to 10% of the normal samples) from the data. This allows to simulate unsupervised settings with an anomaly ratio of up to 10% of entire training set. Rest of the data is used for testing. For MVTec, since there are no anomalous data available for training, we borrow 10% of the anomalies from the test set and swap them with normal samples in the training set.
Hardware Specification Yes Each experimental run is performed on a single V100 GPU.
Software Dependencies No The paper mentions using specific models (e.g., Res Net-18 architecture) and optimizers (Momentum SGD) but does not provide specific version numbers for software libraries or programming languages.
Experiment Setup Yes The same model and hyperparameter configurations are used for SRR with K = 5 classifiers in the ensemble. We set γ as twice the anomaly ratio of training data. For 0% anomaly ratio, we set γ as 0.5. Finally, a Gaussian Density Estimator (GDE) on learned representations is used as the OCC. Optimizer Momentum SGD (momentum= 0.9) Learning rate 0.001 Batch size 64 M L2 weight regularization 0.00003 Random projection dimension 32