Interpretable Image Classification with Adaptive Prototype-based Vision Transformers

We present ProtoViT, a method for interpretable image classification combining deep learning and case-based reasoning. This method classifies an image by comparing it to a set of learned prototypes, providing explanations of the form ``this looks like that.'' In our model, a prototype consists of \textit{parts}, which can deform over irregular geometries to create a better comparison between images. Unlike existing models that rely on Convolutional Neural Network (CNN) backbones and spatially rigid prototypes, our model integrates Vision Transformer (ViT) backbones into prototype based models, while offering spatially deformed prototypes that not only accommodate geometric variations of objects but also provide coherent and clear prototypical feature representations with an adaptive number of prototypical parts. Our experiments show that our model can generally achieve higher performance than the existing prototype based models. Our comprehensive analyses ensure that the prototypes are consistent and the interpretations are faithful.

Paper

References (54)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer Mc2e6/10 · confidence 3/52024-07-03

Summary

The authors present a novel method for interpretable image classification by incorporating a vision transformer (ViT) into the prototypical neural network framework which provides case-based reasoning to neural network based image classifiers. They claim that most existing prototypical methods are convolutional neural network (CNN)-based, and those methods are limited by spatially rigid prototypes, thus failing to handle geometric variations of objects. Existing methods that try to handle this geometric variation either rely on a continuous latent space, which is not compatible with ViTs, or are other prototype-based ViT that fail to provide inherently interpretable explanations. Due to these existing problems, the author present ProtoViT with the following contributions: Incorporates a ViT backbone that can adaptively learn interpretable prototypes that can handle geometric variation of different sizes. They achieve the above with a greedy matching algorithm utilizing an adjacency mask and an adaptive slots mechanism. They give empirical evaluation showing SOTA accuracy and a qualitative analysis showing the faithfulness and coherence of the learned prototype representations.

Strengths

Soundness: The methods are clean and sound with ample ablation experiments, and it appears their approach can perform better than others (marginally) and have interpretable and coherent prototypes. Presentation: The paper was easy to follow with clear claims, ideas and methods. While some of the figures could be a bit cleaner (such as the boundaries and borders in Figure 4), and some notation seemed a bit odd, they communicated their ideas/methods well. Contribution: The paper gives a clean method for utilizing ViTs in the prototypical framework of deeplearning for interpretability and even incorporates existing methods for making prototypes more flexible via utilizing the approach from Deformable ProtoPNet. They incorporate a novel coherence loss that encourages sub prototypes to be similar to each other. In addition, this paper utilizes a greedy matching algorithm with an adaptive mask to learn geometrically local sub-prototypes. This method also allows for an adaptive number of sub-prototypes through the slot pruning mechanism.

Weaknesses

Soundness: The lack of qualitative comparison with other ViT methods. I know the authors state that these other vision transformer methods do not project the learned prototypical features to the closest latent patches, but they still provide explanations. Could more be expanded on this and/or a figure showing this lack of reasoning/inherent interpretability? - This is my biggest concern Presentation: In the section 3.4 for “Optimization of last layers” did you mean “... l-th class prototypes…” with a plural on the prototypes? This was unclear to me Contribution: However due to already preexisting ViT prototype methods (with a lack of comparison to them), the contribution this ViT makes compared to others is unclear.

Questions

Examples in the paper only show good outcomes. What do incorrect predictions look like, and does the explanation (prototypes) show a reason for why the model was incorrect? Do the authors have additional comments about potential information leakage between image patches in the attention step occurring in the activation of the prototypes? I ask this in context of a latent patch containing more information regarding other patches around it contributing to the learned prototype. When we project on this prototype, we look at its spatial position and project to the real image, but it may not tell the whole story. When determining the next sub-prototype, why do you only consider the last sub-prototype when creating the adaptive mask given radius r? Why not all the currently selected sub-prototypes? Wouldn’t you get more coverage and cohesive information if considering them all? Why can’t prototypes be shared across classes? While I still think the prototypes learned are good, I think a limitation that wasn’t stated is the lack of across class prototype sharing. I feel like the coherence loss could limit your overall prototype representation. If all sub-prototypes have to be similar, then wouldn’t that hurt a prototype representation that has distinct parts. I’m thinking of a potential example of three differently colored stripes being next to each other. If the sub-prototypes are each of different colors, wouldn’t they be very different; thus discarded due to this loss? This is even shown in E.3. I know that is where the adaptive mask could combat this problem, but I would like more discussion on this. I’m curious how sensitive your method is to masking / perturbation. For example, if you mask out some or all the sub-prototypes for the top activated area in the test image, can the model still find the prototype elsewhere if it still exists in the image? A particular case I’m interested in is if a prototype is about the blue on a bird’s feathers, so you make that part that the model initially detected. Does the model still pick up the blue elsewhere? Could you expand more on why you “believe that simply adding prototype layers to an architecture without well defined “cases” does not make the new architecture more interpretable.” In particular, what do you mean by ‘well defined cases’?

Rating

6

Confidence

3

Soundness

2

Presentation

3

Contribution

2

Limitations

The method lacks across class prototype sharing. The coherence loss lack discussion on its limitation of making sub-prototypes being similar thus hindering diverse representation which may be important in a prototype. The paper lacks comparison to other interpretable ViT methods

Reviewer Mc2e2024-08-11

Rebuttal Response

Thank you for a detailed rebuttal! I understand that other methods may not project to a latent of a training example, but this doesn't seem like a big step / modification to add to the other methods for comparison. If this was done then having this comparison would have added strength to this paper. I understand that they would just be similar to ProtoPNet if this was added, but they wouldn't be the same. Thank you for emphasizing that other ViT-based methods are unable to handle geometric variations. Thank you for adding parts of this discussion in the paper if accepted. On this point overall, having a unifying comparison between these approaches would have contributed well to this area. Thank you for the perturbation and misclassification examples! Please do add this to the final paper if accepted! I appreciate the response on the coherence loss. I think further analysis of this loss and it's limitations would be desirable. I understand that this is in the latent space, but the concern is still that same. If several parts are semantically different, but their combination is essential for the downstream task, then I still think this loss would restrict that. I understand this loss adds more control which is a reason I added it as a strength, but I still think it comes with limitations (unless there are more analysis that says otherwise). Thank you for addressing some of my other concerns. I've updated my score to reflect the added strength from the rebuttal.

Authorsrebuttal2024-08-11

Thank you for your response! > I understand that other methods may not project to a latent of a training example, but this doesn't seem like a big step / modification to add to the other methods for comparison. If this was done then having this comparison would have added strength to this paper. We agree that adding visualizations from other VIT based models would be very interesting and could directly show the strength of our method. However, since those methods also learn prototypes on the class tokens which have no correspondence to the image as we discussed in the related work, it is unclear how to project and visualize those prototypes even with projection. The main reason that those methods do not perform projection is because they observe a dramatic performance drop. This is explicitly stated on page 11 of ProtoPFormer, where they say: "Our proposed ProtoPFormer does not employ the “push” process for two main reasons. One is that this process causes the performance degradation with ViT backbones, and the other is that our global and local prototypes are the high-level abstraction of associated visual explanations, representing their features based on the whole training set." This drop is because that their method is not able to learn a well clustered latent space, which reviewer **GK9W** agreed on. Thus, it is extremely unclear how those methods produce the visualizations used in their manuscripts. >Thank you for the perturbation and misclassification examples! Please do add this to the final paper if accepted! We will make sure to do so! >If several parts are semantically different, but their combination is essential for the downstream task, then I still think this loss would restrict that. When they are semantically different, it would be captured by other prototypes associated with that class as encouraged by orthogonality loss, which is why we adopt the orthogonality loss in our method. Coherence loss is encouraging the similarity inside a prototype (which consists of some sub-prototypes), and **orthogonality loss is encouraging prototypes to be different from each other (to have diverse representation)**. So, if there are several semantically different parts that are each necessary for classification, a different prototype should learn to identify each one of them. We will make sure to discuss this more in the final manuscript. We hope this would further address your concerns. And thank you again for updating your score and engaging with us! We greatly appreciate your feedback.

Reviewer GK9W5/10 · confidence 5/52024-07-08

Summary

The authors introduce ProtoViT, a model that leverages the Visual Transformer (ViT) architecture and integrates prototypical parts for case-based reasoning. This method is self-explainable, adhering to the rule "this looks like that." A novel aspect of ProtoViT is the use of prototypical parts of varying sizes, utilizing a ViT backbone. The authors claim that these prototypical parts are coherent and inherently interpretable. They evaluate ProtoViT on the CUB and Stanford Cars datasets, using accuracy as the metric for comparison. The training process for the model involves five loss components in addition to cross-entropy.

Strengths

The paper is well-written, with images effectively illustrating the intended concepts. The introduction of the greedy matching algorithm is particularly engaging and holds significant importance for the community. The introduction section is well-crafted, clearly outlining the contributions. Additionally, the computational experiments are thorough and demonstrate comprehensive accuracy.

Weaknesses

The primary concern lies in ensuring that the ViT backbone can maintain prototypical parts that are both local and inherently interpretable. Since ViT uses attention mechanisms that mix information from all patches, there is a risk of confusion for the end user. To address this, I suggest conducting a spatial misalignment benchmark [1] to analyze its influence. Additionally, there are no metrics related to explainability demonstrating whether the model improves interpretability, such as with FunnyBirds [2] or through a user study [3]. [1] Sacha, Mikołaj, et al. "Interpretability benchmark for evaluating spatial misalignment of prototypical parts explanations." Proceedings of the AAAI Conference on Artificial Intelligence. Vol. 38. No. 19. 2024. [2] Hesse, Robin, Simone Schaub-Meyer, and Stefan Roth. "FunnyBirds: A synthetic vision dataset for a part-based analysis of explainable AI methods." Proceedings of the IEEE/CVF International Conference on Computer Vision. 2023. [3] Kim, Sunnie SY, et al. "HIVE: Evaluating the human interpretability of visual explanations." European Conference on Computer Vision. Cham: Springer Nature Switzerland, 2022.

Questions

How are you measuring all the claims, especially coherence and inherently interpretable models? Could you provide a prototype purity metric for PIP-Net? Is it possible to analyze whether ViT can be a backbone for prototypical parts or not, test with spatial misalignment benchmark?

Rating

5

Confidence

5

Soundness

2

Presentation

3

Contribution

3

Limitations

There is no quantification of interpretability, no user-study, nor no reference to XAI benchmarks such as FunnyBirds and spatial misalignment benchmarks.

Reviewer w1UG7/10 · confidence 5/52024-07-11

Summary

This paper presents a novel strategy to learn interpretable visual prototypes for visual transformers, with a good property of offering spatially deformed prototypes. The method also introduce an slot mechanism which can learn an adaptive number of prototypical parts. The proposed are wisely designed for visual transformer architectures.

Strengths

1. It is nice to draw inspiration from the focal similarity when computing the patch features. 2. Different from Deformable ProtoPNet, the proposed method present a new way to accommodate geometric variations of objects. 3. The proposed method is validated on two benchmarks and with extensive ablation studies. 4. In general, the paper is well written.

Weaknesses

1. The first concerns is about the use of deformable prototypes that containing K sub-prototypes, which means the proposed method will have more learnable prototype vectors (K times) than ProtoPNet, TesNet, ProtoPFormer, and so on. These previous approaches only use 1*1 prototypes. Does the performance improvement of the proposed method come from the largely increased number of sub-prototypes? 2. The idea of greedy matching algorithm is similar to the greedy prototype projection, proposed in [1, 2]. The authors are suggested to state their difference of the related works. Also, some important work using prototypes for interpretable image classification should be reviewed and discussed, such as [3, 4, 5]. 3. Regarding the adaptive slots mechanism, it is good for the motivation of to learn an additional indicator to measure the importance of sub-prototypes. From my understanding, the learnable vector v is like a gate, which should be saved as model parameters after training. One potential limitation is such mechanism introduces an extra gate parameter v, compared with previous methods. 4. It not much clear about the adjacency masking. Since the prototypes are not initialized to have the position information, how do choose the patch/feature tokens around the prototypes within r? 5. The authors mention the issue of performance degradation after prototype projection. Does the proposed method still suffer from such issue? What extent will the performance drop? 6. The method has too much loss coefficients, which are selected without detailed tuning procedure. References: [1] Knowledge Distillation to Ensemble Global and Interpretable Prototype-Based Mammogram Classification Models [2] Pixel-grounded prototypical part networks [3] PIP-Net: Patch-Based Intuitive Prototypes for Interpretable Image Classification [4] Learning support and trivial prototypes for interpretable image classification [5] Concept-level debugging of part-prototype network

Questions

N/A

Rating

7

Confidence

5

Soundness

4

Presentation

3

Contribution

4

Limitations

N/A

Reviewer dybc6/10 · confidence 4/52024-07-24

Summary

The paper presents ProtoViT, a method for interpretable image classification. ProtoViT incorporates ViT backbone with deformed prototypes that explains its predictions. ProtoViT consists of three components: (1) a feature encoding layer with a pre-trained ViT backbone, which computes a latent representation of an image; (2) a greedy matching layer which compares the latent representation to learned prototypes; and (3) an evidence layer which aggregates prototype similarity scores into a classification using a fully connected layer. Quantitatively, ProtoViT achieves better performance than previous prototype-based methods. Qualitatively, it identifies meaningful prototypes to explain the prediction.

Strengths

1. The paper is well-written and illustrates the technical details. The qualitative visualization are helpful in understanding the method. Attached appendix provides a lot of meaningful details. 2. The overall scheme of interpretable image classification is nicely presented with small patches!

Weaknesses

1. It is understandable that one goal of interpretable image classification is to tell why something works. However, a bigger goal is to understand why something didn't really work. Is it possible to highlight the examples where the method could not predict the correct class? It would be then interesting to understand the reasons for failure. 2. The method relies on a strong assumption of a solid backbone model that can already achieve a good performance on the given task. This restricts the applicability of the method to limited scenarios where we do not need interpretable image classification at the first place. In such a scenarios, the nearest neighbors obtained using latent feature representation and pixel correspondences may be sufficient. 3. Limitations of the method is not clear from the paper. I could guess at certain places in the method section where things could go wrong. It would be better if the authors could illustrate those points with suitable qualitative analysis. 4. User study is not properly designed and is not statistically significant. Overall: The paper is interesting. In a first reading, everything looks good. But then you start asking yourself questions about the different scenarios where the method will not work (and there are definitely such scenarios as can be judged from quantitative evaluation and user study) and why it will not work. The paper falls short in explaining them with details.

Questions

Please see Weaknesses Section.

Rating

6

Confidence

4

Soundness

3

Presentation

3

Contribution

2

Limitations

Authors did provide limitations but a far-fetched one that does not really talk about the limitations of the current method.

Reviewer GK9W2024-08-09

That is indeed a great response to my review. Thank you for the clarifications and discussion. If your work is accepted, please include values for the purity metric, as it is an established metric for evaluation. Including this will enhance the consistency of the research. The results of the spatial misalignment benchmark are surprising, as I did not expect the ViT to be more robust in terms of explanation quality and faithfulness. Good job! Regarding the user study, I really appreciate the presentation of results from Ma et al. [1], who also tested if users performed better than random. It would be worthwhile to check this for your method as well. After consideration, I am leaning toward increasing my grade, but I still need to digest the information from other reviewers and their discussions. [1] Ma, Chiyu, et al. "This looks like those: Illuminating prototypical concepts using multiple visualizations." Advances in Neural Information Processing Systems 36 (2024).

Authorsrebuttal2024-08-09

Thank you for your response! > If your work is accepted, please include values for the purity metric, as it is an established metric for evaluation. Including this will enhance the consistency of the research. We will make sure to include the purity metric to the final manuscript if the paper is accepted. > Regarding the user study, I really appreciate the presentation of results from Ma et al. [1], who also tested if users performed better than random. It would be worthwhile to check this for your method as well. We agree that this could be an interesting addition to our paper, but it will take a long time to design and complete such a user study. We don't believe that we can finish it before the end of the discussion period. We will add it to the final manuscript if the paper is accepted. Thanks again for engaging with us!

Reviewer w1UG2024-08-11

Thank you for the responses. I am happy with your work. It is great to have an exhaustive discussion about the performance drop caused by prototype projection. I hope such discussion can be included in the main paper, if your work is accepted, since a few prior work noticed this practical issue. Good job!

Authorsrebuttal2024-08-11

Thank you very much for your response! We will make sure to include the discussion of the performance drop caused by projection in the body of the final manuscript. It will be _discussed in the training algorithm section after the discussion of projection starting on page 7_. It is encouraging to know that our detailed revisions have addressed your concerns. Your recognition of our work in terms of contribution, novelty, soundness, and presentation is greatly appreciated. We also appreciate it _if you could consider updating the rating_, as this would kindly reflect your most recent evaluation and kind recognition of the paper and acknowledge our efforts. Thank you again for engaging with us!

Reviewer dybc2024-08-13

Thanks!

Thank you for the rebuttal! I have raised my scores.

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC