Coarse-to-Fine Concept Bottleneck Models

Authors: Konstantinos Panousis, Dino Ienco, Diego Marcos

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

Reproducibility Variable Result LLM Response
Research Type Experimental We experimentally show that CF-CBMs outperform other SOTA approaches classificationwise, while substantially improving interpretation capacity.
Researcher Affiliation Collaboration Konstantinos P. Panousis1,2,5 Dino Ienco1,2,3,4 Diego Marcos1,2 1Inria 2University of Montpellier 3Inrae 4UMR TETIS 5Department of Statistics, AUEB panousis@aueb.gr diego.marcos@inria.fr dino.ienco@inrae.fr
Pseudocode No The paper does not contain explicit pseudocode or algorithm blocks.
Open Source Code Yes Code available at: https://github.com/konpanousis/Coarse-To-Fine-CBMs
Open Datasets Yes We consider three benchmark datasets for evaluating the proposed framework, namely, CUB[21], SUN[23], and Image Net-1k[3].
Dataset Splits No The paper mentions using "validation set" for evaluation but does not explicitly provide the dataset splits (e.g., percentages or counts) for training, validation, and testing.
Hardware Specification Yes We trained our models using a single NVIDIA A5000 GPU with no data parallelization.
Software Dependencies No The paper mentions using Monte Carlo (MC) sampling and the Gumbel-Softmax trick but does not specify software dependencies with version numbers (e.g., Python, PyTorch, specific libraries).
Experiment Setup Yes For our experiments, we set αH = αL = β = 10 4; we select the best performing learning rate among {10 4, 10 3, 5 10 3, 10 2} for the linear classification layer. We set a higher learning rate for W Hs and W Ls (10 ) to facilitate learning of the discovery mechanism. For all our experiments, we use the Adam optimizer without any complicated learning rate annealing schemes. For all our experiments, we split the images into P = 16 patches. For SUN and CUB, we train the model for a maximum of 1000 epochs, while for Image Net, we only train for 100 epochs.