Stochastic Concept Bottleneck Models

Authors: Moritz Vandenhirtz, Sonia Laguna, Ričards Marcinkevičs, Julia Vogt

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We show empirically on synthetic tabular and natural image datasets that our approach improves intervention effectiveness significantly. Notably, we showcase the versatility and usability of SCBMs by examining a setting with CLIP-inferred concepts, alleviating the need for manual concept annotations.
Researcher Affiliation Academia Moritz Vandenhirtz , Sonia Laguna , Riˇcards Marcinkeviˇcs, Julia E. Vogt Department of Computer Science ETH Zurich Switzerland
Pseudocode No The paper does not contain structured pseudocode or algorithm blocks.
Open Source Code Yes 1The code is available here: https://github.com/mvandenhi/SCBM.
Open Datasets Yes As a natural image classification benchmark, we evaluate on the Caltech-UCSD Birds-200-2011 dataset (Wah et al., 2011), comprised of bird photographs from 200 distinct classes. [...] Additionally, we explore another natural image classification task on CIFAR-10 (Krizhevsky et al., 2009) with 10 classes.
Dataset Splits Yes We set N = 50,000, p = 1,500, and C = 100, with a 60%-20%-20% train-validation-test split.
Hardware Specification Yes Resource Usage For the experiments of the main paper, we used a cluster of mostly Ge Force RTX 2080s with 2 CPU workers.
Software Dependencies Yes All methods were implemented using Py Torch (v 2.1.1) (Ansel et al., 2024).
Experiment Setup Yes All models are trained for 150 epochs for the synthetic and 300 epochs for the natural image datasets with the Adam optimizer (Kingma & Ba, 2015) with a learning rate of 10 4 and a batch size of 64.