Pith. sign in

REVIEW 3 major objections 6 minor 50 references

SIDE: Sparse Information Disentanglement for Explainable Artificial Intelligence

T0 review · 3 major / 6 minor · reviewed 2026-08-15 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read SIDE shows that prototype-based explanations of image classifiers can be made sparse — about nine prototypes per ImageNet prediction instead of 768 — without sacrificing accuracy.

desk verdict A solid engineering improvement to prototype-based XAI, but the headline explanation-size reduction is partly an artifact of comparing pruned SIDE to unpruned InfoDisent; deserves review with major revision. read the letter →

arxiv 2507.19321 v1 pith:QHTMQANB submitted 2025-07-25 cs.CV cs.AIcs.LG

Viktar Dubovik , {L}ukasz Struski , Jacek Tabor , Dawid Rymarczyk This is my paper · ORCID
classification cs.CVcs.AIcs.LG
keywords prototypicalpartssparseexplanationsinterpretablemachinelearningmulti-labelclassificationasymmetriclosshardpruningfine-grainedImageNet-scaleinterpretability
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper introduces SIDE, a method for making prototype-based explanations of deep image classifiers drastically more compact. On ImageNet with a SwinV2-S backbone, SIDE reaches 82.6% top-1 accuracy while using on average 8.6 prototypes per prediction, compared with InfoDisent's 81.4% accuracy and 768 prototypes. SIDE achieves this by replacing the softmax head with independent sigmoid activations, training with an asymmetric multi-label loss, and then hard-pruning the class-to-prototype weights, followed by fine-tuning and a calibration stage. The paper argues that this sparse, disentangled prototype space is faithful to model behavior, reporting improved or matched scores on the FunnyBirds interpretability benchmark. If correct, the result shows that compact explanations need not come at the cost of accuracy on large-scale tasks.

What carries the argument

The central mechanism is a sparse, non-negative Scores Sheet: a class-to-prototype weight matrix $W$ whose entries are passed through ReLU so that zeroed weights stay zero, combined with independent sigmoid activations per class and an Asymmetric Loss that suppresses gradient signals from easy negatives. Together these produce emergent sparsity during training without an explicit sparsity penalty. Hard pruning then retains only the top $A$ weights per class, and a one-correct-label-activation (OCLA) regularizer pushes the model to activate exactly one class above threshold $t$, yielding compact explanations. A prototype expansion via a $1\times1$ convolution decouples the number of prototypes from the backbone's native channel count, so the method works on transformers with moderate channel widths.

What would settle it

A user study in which participants are asked to predict or verify a model's decision from SIDE's 8.6-prototype explanations and from InfoDisent's full-prototype explanations; if comprehension is not better for the compact explanations, the central interpretability claim fails.

Watch

Extended reading notes

Core claim

SIDE's central discovery is that prototype-based explanations can be made sparse without sacrificing predictive accuracy by treating classification as a multi-label problem rather than a single-label one. The method disentangles backbone feature maps into channels, upsamples the channel dimension with a $1\times1$ convolution, maps each channel to a prototype activation through the sparse pooling operator, and computes class scores with a non-negative ReLU linear layer followed by sigmoid. Training uses Asymmetric Loss, which naturally drives most prototype-class weights to zero during pretraining; a hard pruning step then keeps only the top $k$ weights per class; fine-tuning recovers accuracy; and a final OCLA calibration stage ensures that usually exactly one class is activated. The outcome is a model whose local explanation for a prediction is a handful of prototypes (8.6 on ImageNet at $A=10$, 6.9 at $A=7$) rather than the full prototype set, while matching or exceeding InfoDisent's accuracy.

Load-bearing premise

The paper's central interpretability claim rests on treating explanation compactness—fewer prototypes per prediction—as a proxy for human understanding; no human study or external benchmark ties the reduced prototype count to actually better comprehension.

Editorial extensions

If this is right

  • On ImageNet with a SwinV2-S backbone, SIDE matches InfoDisent's accuracy (82.6% vs. 81.4%) while reducing the average local explanation from 768 prototypes to 8.6, and to 6.9 with stronger pruning.
  • Across CUB, Stanford Cars, and Stanford Dogs, SIDE keeps accuracy within 0.7 points of InfoDisent while cutting global prototype counts by more than 10x and local counts by more than 300x.
  • The four-stage training schedule recovers accuracy after pruning, and the OCLA calibration stage restores the model's tendency to activate exactly one class, keeping explanations focused.
  • SIDE's explanations outperform InfoDisent's on the FunnyBirds correctness and completeness dimensions, indicating that the sparse prototypes align with the model's actual decision evidence.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The paper does not test whether fewer prototypes actually improve human comprehension; a direct behavioral study is the natural next step.
  • If the sparsity effect comes from the combination of Asymmetric Loss, ReLU activations, and pruning rather than from SIDE-specific wiring, the same recipe should transfer to other prototype-based methods.
  • Because sigmoid class probabilities are independent, SIDE's explanations could extend naturally to multi-label and open-set recognition, where softmax heads tend to mask ambiguity.
  • The OCLA metric itself could serve as a simple, threshold-based interpretability score for any prototype-based model, independent of SIDE's training procedure.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 6 minor

Summary. The paper introduces Sparse Information Disentanglement for Explainability (SIDE), an extension of InfoDisent for prototypical-part explanations. SIDE decouples the prototype count from the backbone feature dimension via a 1x1 convolution, replaces softmax with sigmoid activations and cross-entropy with asymmetric loss, hard-prunes the Scores Sheet to about A prototypes per class, and adds a calibration stage with a One Correct Label Activation (OCLA) regularizer. Experiments on CUB, Stanford Cars, Stanford Dogs, and ImageNet report top-1 accuracy close to InfoDisent while using far fewer prototypes per prediction, and FunnyBirds results are reported for explanation correctness, completeness, and contrastivity. The central claim is that SIDE matches InfoDisent's accuracy while reducing explanation size by over 90%, thereby substantially enhancing the understandability of prototype-based explanations.

Significance. If the size reduction is measured against a matched baseline, the paper would make a useful contribution: it shows that a sigmoid-based, pruned prototype head can maintain accuracy on ImageNet with SwinV2-S (82.6% vs. 81.4% for InfoDisent) while activating only about 8.6 prototypes per image. The paper is strengthened by extensive ablations, experiments across CNN and ViT backbones, and a FunnyBirds evaluation. The main weaknesses are the asymmetric size comparison, the internal inconsistency of the reported size-reduction factors, and the use of a self-defined and directly optimized metric as evidence for interpretability; these need to be fixed before the headline claims can be accepted.

major comments (3)
  1. [§5.1, Tables 1–2] The claim of "reductions of over 10x in Global Size" is not supported by the tables. On CUB with ResNet-50, Global Size drops from 2048 to 310 (6.6x); on DenseNet-121 it drops from 1024 to 266 (3.9x); on ImageNet, SIDE's Global Size (1340 for A=10, 947 for A=7) is larger than InfoDisent's 768. The claim in §5.2 that SIDE explanations are "over 100x smaller" than InfoDisent's is also not supported by Table 2, where 768 divided by 8.6 is approximately 89x. Please re-state the quantitative claims so that they match the reported numbers.
  2. [§5.1–§5.2, Tables 1–2, Figure 1] The Local Size comparison is asymmetric. InfoDisent is assigned a Local Size equal to its total prototype count (2048, 1024, or 768), even though Figure 1 states that for an example ImageNet prediction only 21 of 768 prototypes are active above 0.2. SIDE's Local Size is the output of its own hard-pruning (§4.3) and OCLA thresholding (§4.2) protocol, which by construction keeps about A prototypes per class and forces exactly one class above threshold t. The reported ">300x" Local Size reduction therefore conflates the method's effect with the extraction rule. The authors should apply the same thresholding or top-k extraction to InfoDisent and report matched Local Size values; until then, the headline reduction is not an independent property of SIDE.
  3. [§4.2–§4.3, §5.3] The OCLA metric is defined in Eq. (3), and then the OCLA loss (Eq. 4) is added in the calibration stage to optimize exactly that metric; reporting improved OCLA and smaller Local Size as evidence of interpretability is therefore circular with respect to the design objective. In addition, the FunnyBirds evaluation in §5.3 measures alignment of explanations with model behavior (correctness, completeness, contrastivity), not human understanding, so the abstract's claim that SIDE "substantially enhances the understandability" of prototype-based explanations is not supported by the presented evidence. Please either soften the claim or add a human-subject study that tests whether the compact explanations are actually easier for users to understand.
minor comments (6)
  1. [Abstract, §5.2, Table 2] The text reports 82.8% top-1 accuracy for SIDE (A=10) and a 0.6 pp gap to the 83.4% baseline, but Table 2 shows 82.6% accuracy, which is an 0.8 pp gap. Please reconcile the numbers.
  2. [§4.2, Eq. (3)] The typeset formula for OCLA does not match the prose definition ("exactly one class is activated"). As printed, the expression appears to check only that the ground-truth class is activated. Please rewrite the formula with an explicit logical conjunction, for example \mathbb{1}\left[\left(\sum_j \mathbb{1}(P_{ij} > t)\right) = 1 \wedge P_{iY_i} > t\right].
  3. [§5.3, experimental setup] The word "completeness" is misspelled as "completness" twice in §5.3, and "Densnet-121" appears in the experimental setup text instead of "DenseNet-121."
  4. [Tables 1 and 2] The tables report single-run accuracies with no standard deviations or number of seeds. Given that several comparisons are within a few tenths of a percentage point (e.g., 82.0 vs. 82.6 on CUB DenseNet-121), please report variability across multiple runs.
  5. [Code availability] The paper states "we make code available" but does not provide a URL. Please include a repository link.
  6. [Limitations] The limitations section mentions the multi-stage training procedure; it would also be useful to note the sensitivity of the headline compactness result to the choice of A, t, and the OCLA coefficient λ, since these parameters directly determine the reported Local Size and Global Size.

Circularity Check

2 steps flagged · score 6.0 of 10

The headline >90% explanation-size reduction is largely a by-construction consequence of the hard-pruning hyperparameter A and of an OCLA loss defined to optimize the same metric that defines local explanation size; the accuracy comparison itself is an independent empirical result.

  1. fitted input called prediction [Section 4.3 (Hard Pruning) and Table 2]
    "After pretraining, we apply hard pruning by leaving only top-k scores of W, where k = A∗C and A is desired average number of prototypes activated per class and zeroing-out all the other weights. ... SIDE (A=10) 82.6 1340 8.6"

    The paper's headline 'reducing explanation size by over 90%' uses a Local Size figure that is set by the pruning hyperparameter A: hard pruning retains only A*C prototype-class connections, so the average number of prototypes per class is A by construction. The reported SIDE Local Size (8.6 for A=10, 6.9 for A=7) is thus a direct output of the chosen A, not an emergent property. InfoDisent is not pruned and is assigned Local Size 768 or 2048, so the >90% reduction is an accounting artifact of comparing a pruned-by-construction model with an unpruned baseline.

  2. self definitional [Section 4.2, Eq. (3)-(4); Section 4.3 Calibration]
    "In our SIDE, we define the local explanation for a sample as the set of prototypes associated with all activated classes. ... we introduce OCLA loss — a regularization applied at the final stage of training to ensure that exactly one class probability exceeds threshold t. ... improving OCLA as well as Local Size of the model."

    Local Size is defined through the set of activated classes, and the OCLA loss (Eq. 4) is explicitly built from the same threshold t to suppress all P_ij > t for j ≠ Y_i and push P_i,Y_i above t. Thus exactly one class is activated by construction, which directly shrinks the local explanation and raises the OCLA metric that Eq. (3) defines. Reporting the post-calibration OCLA and Local Size improvements as evidence of interpretability therefore amounts to optimizing the definition itself rather than measuring an independent property.

full rationale

The accuracy results, including ImageNet 82.6% for SIDE versus 81.4% for InfoDisent at the same pruned prototype budget, are not circular: they come from training and evaluation against an external benchmark. The method is also not built on a load-bearing self-citation chain; InfoDisent is a legitimate baseline and base architecture. However, the paper's central compactness claim is substantially enforced by construction. The hard-pruning stage sets the average number of prototypes per class to A, and the calibration-stage OCLA loss directly optimizes the same thresholded activation condition used to define both the OCLA metric and local explanation size. Consequently, the reported reduction from 768 or 2048 prototypes to about 6-9 prototypes per prediction is largely a restatement of the chosen hyperparameters and loss design, not an emergent discovery. The FunnyBirds evaluation measures alignment with model behavior rather than human understanding, so the qualitative claim of 'substantially enhancing understandability' is an interpretive step beyond the measured quantities; this is a validity gap but not itself a circularity. Overall, the partial circularity is concentrated in the explanation-size and OCLA evidence, warranting a score of 6.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The central claim rests on several hand-chosen hyperparameters, especially the pruning budget A and the OCLA calibration settings, which together determine the headline explanation-size figures. The method also inherits InfoDisent's assumptions about frozen features and orthogonal disentanglement. No new physical or ontological entities are introduced.

free parameters (5)
  • Pruning budget A = 5 for fine-grained, 7 and 10 for ImageNet
    Desired average number of prototypes activated per class; sets k = A * C in hard pruning (Section 4.3). This parameter directly determines the reported Local Sizes, so the compactness result is largely a consequence of its chosen value.
  • ASL parameters gamma- and margin m = gamma- in [2,4], m in [0.05,0.2] per Section 3.2; specific values in supplementary
    Asymmetric Loss hyperparameters control gradient attenuation for negatives. The paper says small values are used in pretraining and higher values in fine-tuning, but exact values are deferred to supplementary.
  • OCLA threshold t and calibration coefficient lambda = t = 0.5 default; lambda in {100, 200}
    The threshold defines activation for the OCLA metric and loss. Lambda is selected empirically, with lambda = 200 reported as best trade-off in supplementary.
  • Prototype expansion dimension C' = 1024, 1536, 2048
    The 1x1 convolution expands backbone channels to C'. The paper argues the choice is not critical (Figure 7), but it is still a hyperparameter that affects capacity and was varied in experiments.
  • Weight initialization for Scores Sheet = W ~ N(1.0, 0.1)
    Chosen to bias learning toward activating all prototypes initially (Section 4.3). This initialization influences which connections survive pretraining pruning.
assumptions (4)
  • domain assumption Frozen backbone features contain sufficient discriminative information for classification.
    SIDE keeps the backbone frozen and learns only the head, so the validity of all results depends on the pretrained features being rich enough for the target classes.
  • domain assumption An orthogonal channel-wise transform can disentangle meaningful visual concepts.
    Inherited from InfoDisent (Section 3.1). The orthogonal map U is applied without additional supervision, assuming that rotated channels correspond to semantic prototypes.
  • domain assumption The sparse pooling operator mxpool preserves decision-relevant information.
    The information bottleneck (Section 3.1) reduces each channel to a scalar via max of positive and negative parts, assuming no important evidence is lost in pooling.
  • ad hoc to paper Forcing exactly one class above threshold t via the OCLA loss improves interpretability.
    The OCLA metric and loss (Eq. 3, 4) are introduced in this paper, and the calibration stage optimizes directly for this objective. The assumption that one-active-class explanations are more interpretable is not externally validated.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SIDE: Sparse Information Disentanglement for Explainable Artificial Intelligence." pith.science (2026). https://pith.science/paper/QHTMQANB

@misc{pith2026250719321,
  author       = {Pith},
  title        = {Pith review of: SIDE: Sparse Information Disentanglement for Explainable Artificial Intelligence},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QHTMQANB}},
  note         = {Machine review of arXiv:2507.19321}
}
abstract

Understanding the decisions made by deep neural networks is essential in high-stakes domains such as medical imaging and autonomous driving. Yet, these models often lack transparency, particularly in computer vision. Prototypical-parts-based neural networks have emerged as a promising solution by offering concept-level explanations. However, most are limited to fine-grained classification tasks, with few exceptions such as InfoDisent. InfoDisent extends prototypical models to large-scale datasets like ImageNet, but produces complex explanations. We introduce Sparse Information Disentanglement for Explainability (SIDE), a novel method that improves the interpretability of prototypical parts through a dedicated training and pruning scheme that enforces sparsity. Combined with sigmoid activations in place of softmax, this approach allows SIDE to associate each class with only a small set of relevant prototypes. Extensive experiments show that SIDE matches the accuracy of existing methods while reducing explanation size by over $90\%$, substantially enhancing the understandability of prototype-based explanations.

Figures

Figures reproduced from arXiv: 2507.19321 by the authors.

Figure 1
Figure 1. Comparison of SIDE and InfoDisent prototypical expla [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. SIDE explanation for the ImageNet class Picket Fence [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Main components of the SIDE interpretability method. [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Class probability distributions for an ImageNet Rock [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]
Figure 5
Figure 5. Figure 5: SIDE explanation for three Stanford Dogs classes using ResNet-50. With a hard pruning threshold of [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 6
Figure 6. Figure 6: Results of various interpretability methods on Funny Birds framework with Resnet-50 as backbone. SIDE (pretrain) was evaluated [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 7
Figure 7. Figure 7: Number of active weights in the Scores Sheet of SIDE [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]
Figure 8
Figure 8. Figure 8: Distributions of Scores Sheet active weights for In [PITH_FULL_IMAGE:figures/full_fig_p008_8.png]

Review history (2 revisions) →

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

50 extracted references · 42 canonical work pages

  1. [1]

    Interpretable, not black-box, artificial intelli- gence should be used for embryo selection

    Michael Anis Mihdi Afnan, Yanhe Liu, Vincent Conitzer, Cynthia Rudin, Abhishek Mishra, Julian Savulescu, and Ma- soud Afnan. Interpretable, not black-box, artificial intelli- gence should be used for embryo selection. Human Repro- duction Open, 2021. 2

    work page 2021

  2. [2]

    Towards robust interpretability with self-explaining neural networks

    David Alvarez Melis and Tommi Jaakkola. Towards robust interpretability with self-explaining neural networks. In Ad- vances in Neural Information Processing Systems . Curran Associates, Inc., 2018. 2

    work page 2018

  3. [3]

    A case-based interpretable deep learning model for classification of mass lesions in digital mammography

    Alina Jade Barnett, Fides Regina Schwartz, Chaofan Tao, Chaofan Chen, Yinhao Ren, Joseph Y Lo, and Cynthia Rudin. A case-based interpretable deep learning model for classification of mass lesions in digital mammography. Na- ture Machine Intelligence, 3(12):1061–1070, 2021. 2

    work page 2021

  4. [4]

    Asymmetric loss for multi-label classification, 2021

    Emanuel Ben-Baruch, Tal Ridnik, Nadav Zamir, Asaf Noy, Itamar Friedman, Matan Protter, and Lihi Zelnik-Manor. Asymmetric loss for multi-label classification, 2021. 2, 3

    work page 2021

  5. [5]

    Legal requirements on explainability in machine learning

    Adrien Bibal, Michael Lognoul, Alexandre de Streel, and Benoît Frénay. Legal requirements on explainability in machine learning. Artificial Intelligence and Law , 29(2): 149–169, 2020. 1

    work page 2020

  6. [6]

    B-cos net- works: Alignment is all we need for interpretability

    Moritz Böhle, Mario Fritz, and Bernt Schiele. B-cos net- works: Alignment is all we need for interpretability. In Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10329–10338, 2022. 2

    work page 2022

  7. [7]

    Epic: Explanation of pretrained image classification networks via prototype

    Piotr Borycki, Magdalena Tr˛ edowicz, Szymon Janusz, Jacek Tabor, Przemysław Spurek, Arkadiusz Lewicki, and Łukasz Struski. Epic: Explanation of pretrained image classification networks via prototype. arXiv preprint arXiv:2505.12897 ,

    arXiv

  8. [8]

    Approximating CNNs with bag-of-local-features models works surprisingly well on imagenet

    Wieland Brendel and Matthias Bethge. Approximating CNNs with bag-of-local-features models works surprisingly well on imagenet. In International Conference on Learning Representations, 2019. 2

    work page 2019

Show all 50 references
  1. [9]

    Analysis of explainers of black box deep neural networks for computer vision: A survey, 2019

    Vanessa Buhrmester, David Münch, and Michael Arens. Analysis of explainers of black box deep neural networks for computer vision: A survey, 2019. 1

    2019

  2. [10]

    This looks like that: Deep learning for interpretable image recognition, 2019

    Chaofan Chen, Oscar Li, Chaofan Tao, Alina Jade Barnett, Jonathan Su, and Cynthia Rudin. This looks like that: Deep learning for interpretable image recognition, 2019. 1, 2

    2019

  3. [11]

    De- formable protopnet: An interpretable image classifier using deformable prototypes

    Jon Donnelly, Alina Jade Barnett, and Chaofan Chen. De- formable protopnet: An interpretable image classifier using deformable prototypes. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages 10265–10275, 2022. 2

    2022

  4. [12]

    An image is worth 16x16 words: Transformers for image recognition at scale, 2021

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Syl- vain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition ...

    2021

  5. [13]

    How to probe: Simple yet effective techniques for improving post-hoc explanations

    Siddhartha Gairola, Moritz Böhle, Francesco Locatello, and Bernt Schiele. How to probe: Simple yet effective techniques for improving post-hoc explanations. ICLR, 2025. 2

    2025

  6. [14]

    Explaining deep classification of time- series data with learned prototypes

    Alan H Gee, Diego Garcia-Olano, Joydeep Ghosh, and David Paydarfar. Explaining deep classification of time- series data with learned prototypes. In CEUR workshop pro- ceedings, page 15. NIH Public Access, 2019. 2

    2019

  7. [15]

    Weinberger

    Chuan Guo, Geoff Pleiss, Yu Sun, and Kilian Q. Weinberger. On calibration of modern neural networks, 2017. 4

    2017

  8. [16]

    In- terpretable image recognition with hierarchical prototypes

    Peter Hase, Chaofan Chen, Oscar Li, and Cynthia Rudin. In- terpretable image recognition with hierarchical prototypes. In Proceedings of the AAAI Conference on Human Compu- tation and Crowdsourcing, pages 32–40, 2019. 2

    2019

  9. [17]

    Deep residual learning for image recognition, 2015

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition, 2015. 5

    2015

  10. [18]

    Fun- nybirds: A synthetic vision dataset for a part-based analysis of explainable ai methods

    Robin Hesse, Simone Schaub-Meyer, and Stefan Roth. Fun- nybirds: A synthetic vision dataset for a part-based analysis of explainable ai methods. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3981– 3991, 2023. 6

    2023

  11. [19]

    The inaturalist species classification and de- tection dataset, 2018

    Grant Van Horn, Oisin Mac Aodha, Yang Song, Yin Cui, Chen Sun, Alex Shepard, Hartwig Adam, Pietro Perona, and Serge Belongie. The inaturalist species classification and de- tection dataset, 2018. 6

    2018

  12. [20]

    Weinberger

    Gao Huang, Zhuang Liu, Laurens van der Maaten, and Kil- ian Q. Weinberger. Densely connected convolutional net- works, 2018. 5

    2018

  13. [21]

    Proto2proto: Can you recognize the car, the way i do? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10233–10243, 2022

    Monish Keswani, Sriranjani Ramakrishnan, Nishant Reddy, and Vineeth N Balasubramanian. Proto2proto: Can you recognize the car, the way i do? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10233–10243, 2022. 2

    2022

  14. [22]

    Novel dataset for fine-grained image categorization

    Aditya Khosla, Nityananda Jayadevaprakash, Bangpeng Yao, and Li Fei-Fei. Novel dataset for fine-grained image categorization. In First Workshop on Fine-Grained Visual Categorization, IEEE Conference on Computer Vision and Pattern Recognition, Colorado Springs, CO, 2011. 2, 5

    2011

  15. [23]

    Xpro- tonet: Diagnosis in chest radiography with global and local explanations

    Eunji Kim, Siwon Kim, Minji Seo, and Sungroh Yoon. Xpro- tonet: Diagnosis in chest radiography with global and local explanations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15719– 15728, 2021. 2

    2021

  16. [24]

    3d object representations for fine-grained categorization

    Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In 2013 IEEE International Conference on Computer Vision Workshops, pages 554–561, 2013. 2, 5

    2013

  17. [25]

    Unmasking clever hans predictors and assess- ing what machines really learn

    Sebastian Lapuschkin, Stephan Wäldchen, Alexander Binder, Grégoire Montavon, Wojciech Samek, and Klaus- Robert Müller. Unmasking clever hans predictors and assess- ing what machines really learn. Nature Communications, 10 (1), 2019. 1

    2019

  18. [26]

    Focal loss for dense object detection, 2018

    Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dollár. Focal loss for dense object detection, 2018. 3

    2018

  19. [27]

    Swin transformer v2: Scaling up capacity and resolution, 2022

    Ze Liu, Han Hu, Yutong Lin, Zhuliang Yao, Zhenda Xie, Yixuan Wei, Jia Ning, Yue Cao, Zheng Zhang, Li Dong, Furu Wei, and Baining Guo. Swin transformer v2: Scaling up capacity and resolution, 2022. 3 9

    2022

  20. [28]

    Interpretable image clas- sification with adaptive prototype-based vision transform- ers

    Chiyu Ma, Jon Donnelly, Wenjun Liu, Soroush V osoughi, Cynthia Rudin, and Chaofan Chen. Interpretable image clas- sification with adaptive prototype-based vision transform- ers. Advances in Neural Information Processing Systems , 37:41447–41493, 2024. 2

    2024

  21. [29]

    Person- alized interpretability–interactive alignment of prototypical parts networks

    Tomasz Michalski, Adam Wróbel, Andrea Bontempelli, Jakub Lu ´styk, Mikolaj Kniejski, Stefano Teso, Andrea Passerini, Bartosz Zieli´nski, and Dawid Rymarczyk. Person- alized interpretability–interactive alignment of prototypical parts networks. arXiv preprint arXiv:2506.05533, 2025. 2

    2025 arXiv

  22. [30]

    Interpretable Machine Learning

    Christoph Molnar. Interpretable Machine Learning. 3 edi- tion, 2025. 1

    2025

  23. [31]

    This looks like that, because

    Meike Nauta, Annemarie Jutte, Jesper Provoost, and Christin Seifert. This looks like that, because... explaining prototypes for interpretable image recognition. In Joint European Con- ference on Machine Learning and Knowledge Discovery in Databases, pages 441–456. Springer, 2021. 2

    2021

  24. [32]

    Neural prototype trees for interpretable fine-grained image recogni- tion

    Meike Nauta, Ron Van Bree, and Christin Seifert. Neural prototype trees for interpretable fine-grained image recogni- tion. In Proceedings of the IEEE/CVF conference on com- puter vision and pattern recognition , pages 14933–14943,

  25. [33]

    Lucidppn: Unambigu- ous prototypical parts network for user-centric interpretable computer vision

    Mateusz Pach, Dawid Rymarczyk, Koryna Lewandowska, Jacek Tabor, and Bartosz Zieli ´nski. Lucidppn: Unambigu- ous prototypical parts network for user-centric interpretable computer vision. ICLR, 2025. 2

    2025

  26. [34]

    Stop explaining black box machine learn- ing models for high stakes decisions and use interpretable models instead

    Cynthia Rudin. Stop explaining black box machine learn- ing models for high stakes decisions and use interpretable models instead. Nature machine intelligence, 1(5):206–215,

  27. [35]

    Interpretable machine learning: Fundamental principles and 10 grand challenges,

    Cynthia Rudin, Chaofan Chen, Zhi Chen, Haiyang Huang, Lesia Semenova, and Chudi Zhong. Interpretable machine learning: Fundamental principles and 10 grand challenges,

  28. [36]

    Berg, and Li Fei-Fei

    Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, San- jeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. Imagenet large scale visual recognition challenge,

  29. [37]

    Protopshare: Prototypical parts sharing for simi- larity discovery in interpretable image classification

    Dawid Rymarczyk, Łukasz Struski, Jacek Tabor, and Bartosz Zieli´nski. Protopshare: Prototypical parts sharing for simi- larity discovery in interpretable image classification. In Pro- ceedings of the 27th ACM SIGKDD Conference on Knowl- edge Discovery & Data Mining, pages 1420...

    2021

  30. [38]

    Protomil: Multiple in- stance learning with prototypical parts for fine-grained in- terpretability

    Dawid Rymarczyk, Aneta Kaczy ´nska, Jarosław Kraus, Adam Pardyl, and Bartosz Zieli ´nski. Protomil: Multiple in- stance learning with prototypical parts for fine-grained in- terpretability. In Joint European Conference on Machine Learning and Knowledge Discovery in Databases. ...

  31. [39]

    In- terpretable image classification with differentiable proto- types assignment

    Dawid Rymarczyk, Łukasz Struski, Michał Górszczak, Ko- ryna Lewandowska, Jacek Tabor, and Bartosz Zieli´nski. In- terpretable image classification with differentiable proto- types assignment. In European Conference on Computer Vi- sion, pages 351–368. Springer, 2022. 2

    2022

  32. [40]

    Progrest: Prototypical graph regression soft trees for molec- ular property prediction

    Dawid Rymarczyk, Daniel Dobrowolski, and Tomasz Danel. Progrest: Prototypical graph regression soft trees for molec- ular property prediction. SIAM International Conference on Data Mining, 2023. 2

    2023

  33. [41]

    Icicle: Interpretable class incre- mental continual learning

    Dawid Rymarczyk, Joost van de Weijer, Bartosz Zieli ´nski, and Bartlomiej Twardowski. Icicle: Interpretable class incre- mental continual learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1887– 1898, 2023. 2

    2023

  34. [42]

    Protoseg: Interpretable seman- tic segmentation with prototypical parts

    Mikołaj Sacha, Dawid Rymarczyk, Lukasz Struski, Jacek Tabor, and Bartosz Zieli´nski. Protoseg: Interpretable seman- tic segmentation with prototypical parts. In Winter Confer- ence on Applications of Computer Vision (WACV), 2023. 2

    2023

  35. [43]

    Interpretability benchmark for evaluating spatial misalignment of prototyp- ical parts explanations

    Mikołaj Sacha, Bartosz Jura, Dawid Rymarczyk, Łukasz Struski, Jacek Tabor, and Bartosz Zieli ´nski. Interpretability benchmark for evaluating spatial misalignment of prototyp- ical parts explanations. In Proceedings of the AAAI Confer- ence on Artificial Intelligence, pages 21...

    2024

  36. [44]

    Infodisent: Explainability of image classification mod- els by information disentanglement

    Łukasz Struski, Dawid Rymarczyk, and Jacek Tabor. Infodisent: Explainability of image classification mod- els by information disentanglement. arXiv preprint arXiv:2409.10329, 2024. 1, 2

    2024 arXiv

  37. [45]

    Axiomatic attribution for deep networks, 2017

    Mukund Sundararajan, Ankur Taly, and Qiqi Yan. Axiomatic attribution for deep networks, 2017. 7

    2017

  38. [46]

    This looks like it rather than that: Pro- toknn for similarity-based classifiers

    Yuki Ukai, Tsubasa Hirakawa, Takayoshi Yamashita, and Hironobu Fujiyoshi. This looks like it rather than that: Pro- toknn for similarity-based classifiers. In The Eleventh Inter- national Conference on Learning Representations, 2022. 2

    2022

  39. [47]

    The caltech-ucsd birds-200-2011 dataset, 2023

    Catherine Wah, Steve Branson, Peter Welinder, Pietro Per- ona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset, 2023. 2, 5

    2011

  40. [48]

    Interpretable image recognition by con- structing transparent embedding space

    Jiaqi Wang et al. Interpretable image recognition by con- structing transparent embedding space. In Proceedings of the IEEE/CVF International Conference on Computer Vi- sion, pages 895–904, 2021. 2

    2021

  41. [49]

    Protopformer: Concentrating on prototypical parts in vision transformers for interpretable image recognition, 2022

    Mengqi Xue, Qihan Huang, Haofei Zhang, Lechao Cheng, Jie Song, Minghui Wu, and Mingli Song. Protopformer: Concentrating on prototypical parts in vision transformers for interpretable image recognition, 2022. 1

    2022

  42. [50]

    Protgnn: Towards self-explaining graph neu- ral networks

    Zaixi Zhang, Qi Liu, Hao Wang, Chengqiang Lu, and Cheekong Lee. Protgnn: Towards self-explaining graph neu- ral networks. 2022. 2 10

    2022

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.