Window-Based Distribution Shift Detection for Deep Neural Networks

Authors: Guy Bar-Shalom, Yonatan Geifman, Ran El-Yaniv

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

Reproducibility Variable Result LLM Response
Research Type Experimental In a comprehensive empirical study, we compared our CBD algorithm with the best-performing baselines, including the KS approach of [37].
Researcher Affiliation Collaboration Guy Bar-Shalom Department of Computer Science Technion Israel Institute of Technology guy.b@cs.technion.ac.il Yonatan Geifman Deci.AI yonatan@deci.ai Ran El-Yaniv Department of Computer Science Technion Israel Institute of Technology, Deci.AI rani@cs.technion.ac.il
Pseudocode Yes Algorithm 1: Selection with guaranteed coverage (SGC)
Open Source Code Yes Code is available at https://github.com/Bar SGuy/Window-Based-Distribution-Shift-Detection.
Open Datasets Yes Our experiments are conducted on the Image Net dataset [7], using its validation dataset as proxies for the source distribution, P.
Dataset Splits Yes To train our detectors, we randomly split the Image Net validation data (50,000) into two sets, a detection-training or source set, which is used to fit the detectors3 (49,000) and a validation set (1,000) for applying the shift.
Hardware Specification No The paper mentions conducting experiments and utilizing specific models like Res Net50, Mobile Net V3-S, and Vi T-T, but it does not provide any specific details about the hardware (e.g., GPU model, CPU type, memory) used to run these experiments.
Software Dependencies No The paper mentions 'All experiments were conducted using Py Torch' and refers to 'Sci Py s stats.ttest_1samp implementation [44]', as well as models being 'publicly available in timm s repository [45]', but it does not provide specific version numbers for these software components.
Experiment Setup Yes We evaluated the models using various window sizes, |Wk| {10, 20, 50, 100, 200, 500, 1000}. Our coverage-based detection algorithm applies SGC to Ctarget target coverages uniformly spread between the interval [0.1, 1], excluding the coverage of 1. We set Ctarget = 10, δ = 0.01, and κf(x) = 1 Entropy(x) for all our experiments