Invariant Anomaly Detection under Distribution Shifts: A Causal Perspective

Authors: João Carvalho, Mengtao Zhang, Robin Geyer, Carlos Cotrini, Joachim M Buhmann

NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Through extensive experimental evaluation on both synthetic and real-world tasks, covering a range of six different AD methods, we demonstrated significant improvements in out-of-distribution performance.
Researcher Affiliation Academia João B. S. Carvalho, Mengtao Zhang, Robin Geyer, Carlos Cotrini, Joachim M. Buhmann Institute for Machine Learning Department of Computer Science ETH Zürich {joao.carvalho, mengtao.zhang, robin.geyer, ccarlos, jbuhmann}@inf.eth.ch
Pseudocode No The paper does not contain any structured pseudocode or algorithm blocks.
Open Source Code Yes Code is available at: https://github.com/Joao Carv/invariant-anomaly-detection.
Open Datasets Yes For a realistic anomaly detection scenario, we considered the task of identifying tumorous tissue from images of histological cuts, using the Camelyon17 (Koh et al. [2021], Bandi et al. [2018]) dataset.
Dataset Splits Yes Following the generation of these training environments, we proceeded to develop another pair of environments. These new environments were crafted for the validation data.
Hardware Specification Yes The resources supplied were part of a local custer, and consited of two GPU models: the NVIDIA TITAN RTX and the NVIDIA Tesla V100.
Software Dependencies No The main Python libraries used in our implementation, were Pytorch, which is under a BSD-3 license1, and Pytorch Lightning, which is under Apache 2.0 license2. Methods that were derived from the anomalib library (Akcay et al. [2022]), namely STFPM, reverse distillation, and CFA, were already implemented as a Pytorch Lightning Module, and are all under an Apache 2.0 license3. No specific version numbers for these libraries are provided.
Experiment Setup Yes Our approach consisted of two primary steps. The first involved scaling up two key factors: (a) batch size, and (b) learning rate. Subsequently, we methodically scanned through an array of distinct parameters for each baseline model. These included the backbones Res Net18, Res Net34, Res Net50 and Wide Res Net50, alongside various anomaly scoring methodologies that leverage image-level, density estimation, reconstruction error, and pixel-wise density estimation approaches. An additional aspect of our study was an ablation analysis where the regularization weight was fine-tuned by sweeping through the set of values 0.001, 0.01, 0.1, 1, 10, 100.