COMIX: Compositional Explanations using Prototypes
Sarath Sivaprasad, Dmitry Kangin, Plamen Angelov, Mario Fritz
TL;DR
COMIX introduces a novel method for image classification that decomposes images into concepts and traces them to training data, providing faithful and interpretable explanations with improved fidelity and sparsity.
Contribution
It presents a new approach that decomposes images into concepts and matches them with training data to generate faithful explanations, outperforming existing methods.
Findings
48.82% improvement in C-insertion score on ImageNet
Substantial improvements in fidelity and sparsity metrics
Competitive efficiency with other interpretable architectures
Abstract
Aligning machine representations with human understanding is key to improving interpretability of machine learning (ML) models. When classifying a new image, humans often explain their decisions by decomposing the image into concepts and pointing to corresponding regions in familiar images. Current ML explanation techniques typically either trace decision-making processes to reference prototypes, generate attribution maps highlighting feature importance, or incorporate intermediate bottlenecks designed to align with human-interpretable concepts. The proposed method, named COMIX, classifies an image by decomposing it into regions based on learned concepts and tracing each region to corresponding ones in images from the training dataset, assuring that explanations fully represent the actual decision-making process. We dissect the test image into selected internal representations of a…
Peer Reviews
Decision·ICLR 2025 Conference Withdrawn Submission
- S1: The proposed method is well-motivated, and the relationship between the prototypes and the class-defining features is clear. The requirements for intrinsic interpretability that the authors address are essential components of intrinsic XAI approaches.
- W1: One of my primary concerns pertains to the insufficient details regarding the essential computations involved in the proposed method, which necessitate considerable computational resources. In particular, additional clarification is needed on the computation of mutual information maximization as described in Equation 8. Specifically, how are p(F_j) and p(l(F_j) = c) calculated? Does this process necessitate traversing every row of W_{1 \to L}(d, \theta)? This indeed involves significant co
- Originality: The paper proposed an elaborate scheme to turn a B-cos network into a model with a capability to perform case-based reasoning (using k-nearest neighbors). - Quality: The proposed method did not significantly degrade the classification performance. - Clarity: The introduction is well-written, and the paper is well-motivated. - Significance: Interpretability is an important topic.
- Originality: The proposed form of interpretability ("this part of the test image looks like that part of a training image") has been explored in prior work (e.g., ProtoPNet). There is no novelty here. - Quality: The proposed method constructs a model that is not trainable end-to-end. Also, the proposed method is biased toward the pseudo-label predicted by the B-cos network, since the selection of class-defining features are based on the predicted pseudo-label. - Quality: The accuracy of COMiX
- Qualitative results are strong and convincing - Modeling / approach is simple - Robust use in different models / architectures - Robust ability across datasets - Creative use with B-cos networks and label aggregation - Sufficient set of quantitative evaluation metrics - Hyperparameter analysis / ablation studies present - Better sparsity than ViT baseline and best insertion scores w/ competitive deletion scores.
- The interpretability framework novelty isn't significantly more compared to ProtoPNet (Seems like another ‘this looks like that’ explanation just reframed) - Much of the framework is similar to ProtoPNet with the exception of using pretrained features as concepts (as opposed to specialized vectors), a b-cos backbone, and KNN based prediction on feature similarity. - Presentation is unclear at times: - Motivation/need for sufficiency is unclear - Notation seems convoluted - Confused
The paper explores the important field of building interpretable models. The core idea of the paper is presented clearly, with good figures 1 and 5. Combining b-cos with prototypes is a novel combination. The method is evaluated on a sufficient number of datasets and across various architectures.
The paper is missing the discussion of several competitors, e.g.:ProtoPool, Pip-Net, Q-SENN (citations at the bottom); All of these focus on sparsity and Q-SENN and ProtoPool additionally aim for object-level prototypes or features, not restricting the prototypes to patches. They further achieve better results in e.g accuracy and sparsity, which leads to global interpretability. Notably, ProtoPool also uses exact training images as prototype for matching. Thus, the novelty of the proposed method
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsScientific Computing and Data Management
MethodsALIGN · High-Order Consensuses