Multi-dimensional concept discovery (MCD): A unifying framework with completeness guarantees
Authors: Johanna Vielhaben, Stefan Bluecher, Nils Strodthoff
TMLR 2023 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We empirically demonstrate the superiority of MCD against more constrained concept definitions. We carry out our experiments on Image Net (Deng et al., 2009). First, we provide a concrete example for an MCD explanation and showcase its completeness relation introduced in Section 2.3 (benefit 3). To this end, Figure 3 shows an MCD-SSC explanation of a Res Net50v2 prediction for a sample of the police van class in Image Net. We compare MCD with sparse subspace clustering (MCD-SSC), MCD with alternative clustering, and previous methods listed in Table 1 in terms of (1) faithfulness (benefit 1) and (2) conciseness (benefit 2) of the explanations. |
| Researcher Affiliation | Academia | Johanna Vielhaben EMAIL Explainable Artificial Intelligence Group Fraunhofer Heinrich-Hertz-Institute Stefan Blüecher EMAIL Machine Learning Group TU Berlin Nils Strodthoff nils.strodthoff@uol.de Division AI4Health Oldenburg University |
| Pseudocode | No | The paper describes the methodology and algorithms in prose and mathematical equations (e.g., Section 2 and Appendix A). There are no distinct sections or figures explicitly labeled as 'Pseudocode' or 'Algorithm' blocks. |
| Open Source Code | Yes | Code to reproduce our experiments is publicly available at https://github.com/jvielhaben/MCD-XAI. |
| Open Datasets | Yes | We carry out our experiments on Image Net (Deng et al., 2009). We base all our experiments on images from a diverse selection of ten Image Net classes, which roughly align with CIFAR10 classes |
| Dataset Splits | Yes | We carry out our experiments on Image Net (Deng et al., 2009). We base all our experiments on images from a diverse selection of ten Image Net classes, which roughly align with CIFAR10 classes. for each concept l, we sort test set samples by the maximum activation maxxy|φβ,l x,y| and choose the top-k samples as concept prototypes. |
| Hardware Specification | No | The paper mentions model architectures such as Res Net models and Swin Vision Transformer, but does not provide specific details about the CPU, GPU, or other hardware used to run the experiments. |
| Software Dependencies | No | The paper mentions using 'torchvision' and 'timm' for weights of Res Net and Swin-T models, but does not specify version numbers for these software components or any other libraries. |
| Experiment Setup | Yes | For all methods within the MCD framework, we fix the number of concepts in a class-dependent way such that we reach a completeness score of η = 0.5. In all our experiments, we fix the hyperparameter γ, which balances sparsity vs. robustness, to γ = 10. |