Zero-Shot Anomaly Detection via Batch Normalization
Authors: Aodong Li, Chen Qiu, Marius Kloft, Padhraic Smyth, Maja Rudolph, Stephan Mandt
NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our empirical results demonstrate the first zero-shot AD results for tabular data and outperform existing methods in zero-shot anomaly detection and segmentation on image data from specialized domains. |
| Researcher Affiliation | Collaboration | Aodong Li UC Irvine Chen Qiu Bosch Center for AI Marius Kloft TU Kaiserslautern Padhraic Smyth UC Irvine Maja Rudolph Bosch Center for AI Stephan Mandt UC Irvine |
| Pseudocode | Yes | The training pseudo-code is in Alg. 1 of Supp. C. |
| Open Source Code | Yes | Code is at https://github.com/aodongli/zero-shot-ad-via-batch-norm |
| Open Datasets | Yes | We study four image datasets: CIFAR100-C [29], Organ A [81] (and MNIST [42], and Omniglot [41] in Supp. I.4)... We evaluate all methods on two real-world tabular AD datasets Anoshift [17] and Malware [34]. |
| Dataset Splits | Yes | We search the hyperparameters on a validation set split from the training set, after which we integrate the validation set into the training set and train the model on that. Then we test the model on the test set. |
| Hardware Specification | No | The acknowledgements mention 'GPU clusters on which the experiments have been performed', but no specific models, brands, or detailed specifications of the GPUs or other hardware components are provided. |
| Software Dependencies | No | The paper describes implementation details and general frameworks like PyTorch or Adam optimizer, but it does not specify versions for programming languages, libraries, or other software dependencies. |
| Experiment Setup | Yes | We set π = 0.8 in Eq. (6)... For the learning rate, we search values 0.1, 0.01, 0.001, 0.0001, and 0.00001... We also search the mini-batch size B (30, 60) and the number of sub-sampled tasks M (16, 32, 64)... We train the model 6,000 iterations on CIFAR100 data, 10,000 iterations on Omiglot, and 2,000 iterations on MNIST and Organ A. |