Causally motivated multi-shortcut identification and removal

Authors: Jiayun Zheng, Maggie Makar

NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We show both theoretically and empirically that our approach is able to identify a sufficient set of shortcuts leading to more efficient predictors in finite samples. (4) We empirically validate our theoretical findings using a semi-simulated benchmark and a medical task, showing our approach has favorable inand out-of-distribution generalization properties.
Researcher Affiliation Academia Jiayun Zheng Computer Science and Engineering University of Michigan, Ann Arbor Maggie Makar Computer Science and Engineering University of Michigan, Ann Arbor
Pseudocode Yes A full description of the shortcut identification procedure is included in the appendix, section C, procedure 1. Pseudocode for our full approach is included in the appendix section C.
Open Source Code Yes Our code is available on https://github.com/mymakar/cm_multishortcut_id_removal. (a) Did you include the code, data, and instructions needed to reproduce the main experimental results (either in the supplemental material or as a URL)? [Yes] The github link that contains code and instructions is listed in section 6.
Open Datasets Yes We follow Sagawa et al. [36] by constructing a semi-synthetic waterbirds dataset where the task is to predict Y... We combine images of water and land birds extracted from the Caltech-UCSD Birds-200-2011 (CUB) dataset [42] with water and land background extracted from the Places dataset [44]. We use a publicly available dataset made available by Eye PACS, LLC [11]4. (a) If your work uses existing assets, did you cite the creators? [Yes] The creators of all three datasets used in the paper (CUB, Places and DR) are cited in the experiments section. (b) Did you mention the license of the assets? [Yes] We mention in section 6 that all datasets are publicly available.
Dataset Splits Yes We split the training data into two sub samples D1 and D2. Cross-validation. The objective function in (4) depends on two hyperparameters... Letting Dvalid denote a held out validation set, φvalid denote {φ(xi)}i2Dvalid, and similarly define b Vpvalid, our cross validation procedure proceeds as follows.
Hardware Specification Yes (d) Did you include the total amount of compute and the type of resources used (e.g., type of GPUs, internal cluster, or cloud provider)? [Yes] In the appendix.
Software Dependencies No All models in this paper are implemented in Tensor Flow [1]. While TensorFlow is mentioned, no specific version number for it or other software dependencies is provided in the main text to ensure reproducibility.
Experiment Setup Yes Additional details about training are included in the appendix. (b) Did you specify all the training details (e.g., data splits, hyperparameters, how they were chosen)? [Yes] In section 6 and the appendix