Learning Concept Credible Models for Mitigating Shortcuts
Authors: Jiaxuan Wang, Sarah Jabbour, Maggie Makar, Michael Sjoding, Jenna Wiens
NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | 4 Experiments & results In this section, we verify CCM s robustness to spurious correlations on three classification tasks using publicly available datasets. ... Applied to two real-world datasets, we demonstrate that both approaches can successfully mitigate shortcut learning. |
| Researcher Affiliation | Academia | 1Division of Computer Science & Engineering, University of Michigan, Ann Arbor, MI, USA 2Division of Pulmonary and Critical Care, Michigan Medicine, Ann Arbor, MI, USA |
| Pseudocode | No | The paper describes the proposed approaches (CCM RES and CCM EYE) with mathematical formulations, but it does not include a dedicated block labeled 'Pseudocode' or 'Algorithm' with structured steps. |
| Open Source Code | Yes | Code is available at https://gitlab.eecs.umich.edu/mld3/Concept Credible Model. Our code for all experiments is available at https://gitlab.eecs.umich.edu/mld3/Concept Credible Model. |
| Open Datasets | Yes | The Caltech-UCSD Birds-200-2011 dataset (CUB) consists of 11, 788 images of birds [30]... The MIMIC-CXR dataset [32, 33] consists of chest X-rays and corresponding radiology reports. |
| Dataset Splits | Yes | We divide the training set predefined in the CUB dataset into train and validation set with a 80/20 random split, and use the predefined test set for evaluation. ... We divided the chest X-ray datasets into train, validation, and test sets with a 64/16/20 random split. |
| Hardware Specification | No | The paper does not specify any hardware details such as GPU models, CPU types, or memory used for running the experiments. The authors acknowledge this limitation in the checklist: 'Did you include the total amount of compute and the type of resources used (e.g., type of GPUs, internal cluster, or cloud provider)? [No]' |
| Software Dependencies | No | The paper mentions using an 'Inception V3 architecture' and optimization with 'SGD', but it does not provide specific version numbers for any software dependencies like deep learning frameworks (e.g., PyTorch, TensorFlow) or other libraries. |
| Experiment Setup | Yes | All methods are trained on the training set of this biased dataset using SGD with learning rate of 0.01, momentum of 0.9, and batch size of 32. We apply 10 4 weight decay to each model and decay the learning rate every 15 epochs. For CCM EYE, we tune λ in the range of [10 2, 10 3, 10 4, 10 5, 10 6]. |