Addressing Leakage in Concept Bottleneck Models

Authors: Marton Havasi, Sonali Parbhoo, Finale Doshi-Velez

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

Reproducibility Variable Result LLM Response
Research Type Experimental We present empirical results showcasing the efficacy of our proposed modifications to concept bottleneck models (CBM).
Researcher Affiliation Academia Marton Havasi School of Engineering and Applied Sciences Harvard University mhavasi@seas.harvard.edu Sonali Parbhoo Department of Electrical Engineering Imperial College London s.parbhoo@imperial.ac.uk Finale Doshi-Velez School of Engineering and Applied Sciences Harvard University finale@seas.harvard.edu
Pseudocode Yes Psudocode for the interventions is shown in Appendix D.
Open Source Code Yes Our code is available at https://github.com/dtak/addressing-leakage.
Open Datasets Yes MIMIC-III EWS (Johnson et al., 2016) Caltech-UCSD Birds 2011 (Wah et al., 2011)
Dataset Splits Yes The dataset contains records from 17,289 patients over a combined N=796,250 time steps (split into 530,802 training and 265,448 test examples, while ensuring that no patient appears in both sets). This dataset contains N=11,788 (5,994 training and 5,794 test) images of 200 bird species native to North America
Hardware Specification Yes The times are recorded on a V100 GPU.
Software Dependencies No The paper mentions models and optimizers like Inception v3 network and Adam, but does not specify software dependencies with version numbers (e.g., PyTorch 1.x, TensorFlow 2.x, CUDA version).
Experiment Setup Yes For MIMIC-III EWS, the concept predictor is a two-layer feed-forward neural network with a hidden layer of size 100, and the label predictor is a two-layer network with hidden layer of size 50. In the autoregressive case, a small, two-layer network (hidden layer size 20) predicts each concept... We use M = 200 Monte-Carlo samples for prediction. For training hyperparameters, see Appendix C.