On the Analysis and Reproduction of "Post-hoc Concept Bottleneck Models" with an Extension to the Audio Domain
Abstract: Although deep neural networks are powerful tools, they are yet considered "black boxes". With the proliferation of AI models, the need for their interpretability has increased. One way to improve the interpretability of deep neural networks is to understand their decisions in terms of human-understandable concepts. Concept Bottleneck Models (CBMs) aim to achieve this goal by using embedding representations of concepts into the model, providing explainability into the decisions that a network makes. However, CBMs have various limitations concerning training efficiency and task applicability. The authors of the paper Post-hoc Concept Bottleneck Models (PCBMs) provide a novel approach to creating CBMs in a more efficient and generalizable way. In this paper, we evaluate their claims, namely, that PCBMs can be trained using any pre-trained neural network and that PCBMs offer interpretability without sacrificing significant performance. To do so, we not only attempted to reproduce the original paper results but also extended the approach into the audio domain. Our results show good alignment with the original paper but further analysis revealed some problems PCBMs may have, namely, challenges in getting a suitable list of relevant human-understandable concepts for a given task, and potential misalignment between concept encoders and input feature encoders. The code for our paper can be found at https://anonymous.4open.science/r/-354E/
Submission Length: Regular submission (no more than 12 pages of main content)
Changes Since Last Submission: Apart from minor improvements in language and spelling, we incorporated significant improvements pointed out by the reviewers during the rebuttal period of this submission, including improvements to the details and descriptions of results, new results on certain problematic datasets, and a clearer, updated methodology.
We thank the reviewers for their inputs.
Assigned Action Editor: ~Ellen_Vitercik1
Submission Number: 2244
Loading