Transfer-Based Semantic Anomaly Detection
Authors: Lucas Deecke, Lukas Ruff, Robert A. Vandermeulen, Hakan Bilen
ICML 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We rigorously verify our hypothesis in controlled trials that utilize intervention, and show that it gives rise to surprisingly effective auxiliary objectives that outperform previous AD paradigms. Our experiments show that such strategies provide very powerful methods for AD that outperform previous approaches in the deep AD literature on a set of common benchmarks (Section 5). |
| Researcher Affiliation | Collaboration | 1University of Edinburgh 2Aignostics (majority of work done while with TU Berlin) 3ML Group, TU Berlin. |
| Pseudocode | No | The paper does not contain any structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | Code available at https://github.com/VICO-UoE/TransferAD. |
| Open Datasets | Yes | For the data we make use of a standard AD benchmark (Ruff et al., 2018): single classes from CIFAR-10 (e.g. dogs) constitute the normal class... transferring features from semantic tasks such as ILSVRC image classification (Deng et al., 2009)... high-resolution, realistic datasets such as the recently released MPI3D (Gondal et al., 2019)... In the CIFAR-10 and STL-10 semantic AD benchmarks... For experiments on CIFAR-10 (Krizhevsky & Hinton, 2009)... We use the same architecture for STL-10 (Coates et al., 2011). |
| Dataset Splits | No | The paper mentions training and test sets but does not explicitly describe a validation set split or its proportion. For example, 'single classes from CIFAR-10 (e.g. dogs) constitute the normal class, and the one-class model g is learned over all embedded training examples of this class. At test time, we measure whether the two-stage model can successfully identify the appearance of the remaining object classes (cats, deers, etc.) as anomalous.' |
| Hardware Specification | No | The paper does not provide specific hardware details (e.g., GPU/CPU models, memory specifications) used for running the experiments. |
| Software Dependencies | No | The paper states 'All experiments have been implemented with Py Torch (Paszke et al., 2019).' However, it does not provide a specific version number for PyTorch or any other software dependencies. |
| Experiment Setup | Yes | The underlying model for DSVDD, SAD, ADRA, ADIB is the exact same Res Net26, optimized via stochastic gradient descent (momentum parameter of 0.9, weight decay of 10 4) for a total of 100 epochs, with learning rate reductions by 1/10 after 60 and 80 epochs. The batch size is fixed to 128, and we only use standard augmentations. For an explicit inductive bias in ADIB, we scale the regularization term Ω(θ) with α=10 2, as recommended by Li et al. (2018). As before, for ADIB we set α=10 2 following the suggestion of Li et al. (2018); in elastic weight consolidation (EWC) we set the Fisher multiplier to 400, as recommended by Kirkpatrick et al. (2017). |