Post-hoc Concept Bottleneck Models

Authors: Mert Yuksekgonul, Maggie Wang, James Zou

ICLR 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate the PCBM and PCBM-h in challenging image classification and medical settings, demonstrating several use cases for PCBMs. We further address practical concerns and show that PCBMs can be used without a loss in the original model performance. We used the following datasets to systematically evaluate the PCBM and PCBM-h: CIFAR10, CIFAR100 (Krizhevsky et al., 2009) ... In Table 1, we report results over these five datasets.
Researcher Affiliation Academia Mert Yuksekgonul, Maggie Wang, James Zou Stanford University {merty,maggiewang,jamesz}@stanford.edu
Pseudocode No No explicit pseudocode or algorithm blocks are provided. The methodology is described in prose and mathematical equations (e.g., Equation 1 and 2).
Open Source Code Yes The code for our paper can be found in https://github.com/mertyg/post-hoc-cbm.
Open Datasets Yes We used the following datasets to systematically evaluate the PCBM and PCBM-h: CIFAR10, CIFAR100 (Krizhevsky et al., 2009)... CUB (Wah et al., 2011)... HAM10000 (Tschandl et al., 2018)... SIIM-ISIC (Rotemberg et al., 2021).
Dataset Splits Yes We tune the regularization strength on a subset of the training set, that is kept as a validation set.
Hardware Specification Yes We trained all our models on a single NVIDIA-Titan Xp gpu.
Software Dependencies No PCBMs are fitted using scikit-learn s SGDClassifier class, with 5000 maximum steps. Hybrid parts are trained with Py Torch, where we used Adam as the optimizer with 0.01 learning rate, with 0.01 L2 regularization on the residual classifier weights, and trained for 10 epochs.
Experiment Setup Yes Hyperparameters: In all of our experiments Elastic Net sparsity ratio parameter was α = 0.99. We trained all our models on a single NVIDIA-Titan Xp gpu. All of the models were trained for a total number of 10 epochs. We tune the regularization strength on a subset of the training set, that is kept as a validation set. PCBMs are fitted using scikit-learn s SGDClassifier class, with 5000 maximum steps. Hybrid parts are trained with Py Torch, where we used Adam as the optimizer with 0.01 learning rate, with 0.01 L2 regularization on the residual classifier weights, and trained for 10 epochs.