Learning and Evaluating Representations for Deep One-Class Classification
Authors: Kihyuk Sohn, Chun-Liang Li, Jinsung Yoon, Minho Jin, Tomas Pfister
ICLR 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | 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. |
| Researcher Affiliation | Industry | Kihyuk Sohn , Chun-Liang Li , Jinsung Yoon, Minho Jin & Tomas Pfister Google Cloud AI {kihyuks,chunliang,jinsungyoon,minhojin,tpfister}@google.com |
| Pseudocode | No | The paper describes its methods and algorithms in prose and mathematical equations but does not include any explicit pseudocode blocks or algorithms. |
| Open Source Code | Yes | The code is available at https: //github.com/google-research/deep_representation_one_class. |
| Open Datasets | Yes | We evaluate on one-class classification benchmarks, including CIFAR-10, CIFAR1004 [27], Fashion-MNIST [28], and Cat-vs-Dog [29]. We further propose a new protocol using Celeb A eyeglasses dataset [30]... we consider the defect detection benchmark MVTec [31]. |
| Dataset Splits | Yes | Other hyperparameters, such as learning rate or the MLP projection head depth, are cross-validated using small labeled data... To this end, we use 10% of inlier (500) and the same number of outlier examples of CIFAR-10 for hyperparameter selection, and use the same set of hyperparameters to test methods on other datasets, which could demonstrate the algorithm robustness. |
| Hardware Specification | No | The paper mentions training models and using TensorFlow but does not specify any particular CPU, GPU, or TPU models, nor specific computational resources like cloud instances or clusters. |
| Software Dependencies | No | We use scikit-learn [74] implementation of OC-SVMs... We use scikit-learn implementation of Gaussian mixture model... Finally, all experiments are conducted using Tensor Flow [75]. The specific version numbers for scikit-learn and TensorFlow are not provided. |
| Experiment Setup | Yes | Unless otherwise stated, models are trained for 2048 epochs with momentum (0.9) SGD and a single cycle of cosine learning rate decay [73]. L2 weight regularization with coefficient of 0.0003 is applied... To this end, we train all models across all datasets using the same hyperparameter configurations, such as learning rate of 0.01, projection head of depth 8 ([512 8, 128]), temperature τ of 0.2, or batch size of 32. |