B-cosification: Transforming Deep Neural Networks to be Inherently Interpretable

Authors: Shreyash Arya, Sukrut Rao, Moritz Böhle, Bernt Schiele

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

Reproducibility Variable Result LLM Response
Research Type Experimental We perform a thorough study of design choices to perform this conversion, both for convolutional neural networks and vision transformers. We find that B-cosification can yield models that are on par with B-cos models trained from scratch in terms of interpretability, while often outperforming them in terms of classification performance at a fraction of the training cost.
Researcher Affiliation Collaboration 1Max Planck Institute for Informatics, Saarland Informatics Campus, Saarbrücken, Germany 2RTG Neuroexplicit Models of Language, Vision, and Action, Saarbrücken, Germany 3 Kyutai, Paris, France
Pseudocode No The paper describes architectural modifications and strategies but does not provide any formal pseudocode or algorithm blocks.
Open Source Code Yes We release our code and pre-trained model weights at https://github.com/shrebox/B-cosification.
Open Datasets Yes We B-cosify models from Torchvision [54] supervised on Image Net [16].
Dataset Splits Yes Table 4 reports the top-1 classification accuracy on the Image Net validation set...
Hardware Specification Yes We use NVIDIA A100-SXM4-40GB and Quadro RTX 8000 GPUs from internal cluster.
Software Dependencies No While software like PyTorch, Captum, and Adam W are mentioned, specific version numbers (e.g., "PyTorch 1.9") are not provided for any of the software components.
Experiment Setup Yes We optimize using Adam W [26] with cosine scheduling and train for 90 epochs