pith. sign in

arxiv: 2508.14255 · v2 · submitted 2025-08-19 · 💻 cs.LG

Graph Concept Bottleneck Models

Haotian Xu , Tsui-Wei Weng , Lam M. Nguyen , Tengfei Ma This is my paper

Pith reviewed 2026-05-18 21:57 UTC · model grok-4.3

classification 💻 cs.LG
keywords concept bottleneck modelslatent concept graphsmodel interpretabilityimage classificationconcept interventionsconcept correlations
0
0 comments X

The pith

Graph Concept Bottleneck Models capture correlations among concepts using latent graphs to improve classification and interventions while keeping interpretability.

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

Standard Concept Bottleneck Models assume concepts are independent, yet real concepts usually correlate so that changing one affects others. This paper introduces GraphCBMs that add latent concept graphs to standard CBMs to represent those relationships explicitly. The resulting models show higher accuracy on image classification benchmarks, supply extra structural information about concepts, and support more effective interventions because adjustments respect interdependencies. The gains hold across different training procedures and network architectures.

Core claim

GraphCBMs integrate latent concept graphs into Concept Bottleneck Models to model hidden correlations between concepts. This structure yields better image classification performance, richer concept-level interpretability, and stronger intervention results compared with standard CBMs that treat concepts as isolated.

What carries the argument

Latent concept graphs that encode the intrinsic correlations and influence relationships among concepts.

If this is right

  • GraphCBMs achieve higher accuracy than standard CBMs on real-world image classification tasks.
  • The models supply additional concept structure information that improves interpretability.
  • Interventions become more effective because the graph accounts for how one concept change influences others.
  • Performance stays stable when training methods or model architectures vary.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same graph-construction step could be tested on tasks beyond images, such as medical diagnosis where symptoms are interdependent.
  • If the approach generalizes, other intermediate-representation models might gain from explicit relationship graphs rather than assuming independence.
  • Datasets with explicitly annotated concept correlations would provide a direct test of whether the graph component drives the reported gains.

Load-bearing premise

Concepts in these models possess an intrinsic structure in which they are correlated, so that changing one concept necessarily affects its related concepts.

What would settle it

An experiment in which removing the latent graph or forcing concepts to be independent produces no drop in intervention effectiveness or classification accuracy.

Figures

Figures reproduced from arXiv: 2508.14255 by Haotian Xu, Lam M. Nguyen, Tengfei Ma, Tsui-Wei Weng.

Figure 1
Figure 1. Figure 1: Overview of Graph CBM. ① Extract embeddings via pretrained encoders and initialize the graph structure. ② Update the concept embeddings and concept activations through message passing. ③ Use different granularity contrastive losses to qualify the latent graph and make final predictions. used to predict the final label: yˆi = f2(ci). Each element c j i (1 ≤ j ≤ k) reflects the relevance between image vi and… view at source ↗
Figure 2
Figure 2. Figure 2: We compare our models with current SOTA results for corresponding training and dataset settings, i.e., [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: This figure shows the effectiveness of using latent graphs when intervening concepts. Subfigure A compares [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Top: Illustrating part of the learned latent graph under the label-free setting on CUB, alongside representative connected concept pairs. Bottom: G-CBM and G-CEM capture similar and trustworthy concept interactions for the ChestXpert dataset. In this paper, we introduced Graph CBMs, a simple yet effective framework that incorporates graph structures to model concept correlations and enhance interpretabilit… view at source ↗
Figure 5
Figure 5. Figure 5: T-SNE visualization of concept activations distribution. [PITH_FULL_IMAGE:figures/full_fig_p015_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Attacking different types of concepts (nodes) in the CUB dataset can result in different scales of performance [PITH_FULL_IMAGE:figures/full_fig_p015_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: In a large-scale dataset, having latent graphs can still promote models to interact with intervention at different [PITH_FULL_IMAGE:figures/full_fig_p017_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Adding a latent graph to the CDM can significantly improve model performance. [PITH_FULL_IMAGE:figures/full_fig_p017_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Both graphs are G-CBM concept graphs for the CUB dataset. Right: original concept graph. Left: salient [PITH_FULL_IMAGE:figures/full_fig_p018_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Both graphs are G-CBM concept graphs for the ChestXpert dataset. Right: original concept graph. Left: [PITH_FULL_IMAGE:figures/full_fig_p018_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Comparison among models when full interventions. [PITH_FULL_IMAGE:figures/full_fig_p019_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Performance changes as we set different β value to control the latent graph complexity. L Visualization of Learnable Graph 20 [PITH_FULL_IMAGE:figures/full_fig_p020_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: The overview of the CUB’s concept graph in label-free settings. [PITH_FULL_IMAGE:figures/full_fig_p021_13.png] view at source ↗
read the original abstract

Concept Bottleneck Models (CBMs) provide explicit interpretations for deep neural networks through concepts and allow intervention with concepts to adjust final predictions. Existing CBMs assume concepts are conditionally independent given labels and isolated from each other, ignoring the hidden relationships among concepts. However, the set of concepts in CBMs often has an intrinsic structure where concepts are generally correlated: changing one concept will inherently impact its related concepts. To mitigate this limitation, we propose GraphCBMs: a new variant of CBM that facilitates concept relationships by constructing latent concept graphs, which can be combined with CBMs to enhance model performance while retaining their interpretability. Our experiment results on real-world image classification tasks demonstrate Graph CBMs offer the following benefits: (1) superior in image classification tasks while providing more concept structure information for interpretability; (2) able to utilize latent concept graphs for more effective interventions; and (3) robust in performance across different training and architecture settings.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The manuscript proposes Graph Concept Bottleneck Models (GraphCBMs) as an extension of standard Concept Bottleneck Models. It constructs latent concept graphs to capture intrinsic correlations among concepts (which standard CBMs treat as conditionally independent), and claims that the resulting models achieve superior accuracy on image classification tasks, support more effective concept interventions, and remain robust across training regimes and architectures while preserving interpretability.

Significance. If the central claims hold after capacity-controlled validation, the work would meaningfully advance interpretable deep learning by adding an explicit relational inductive bias to concept bottlenecks. This could improve intervention reliability in safety-critical domains where concept dependencies are known to exist.

major comments (2)
  1. [Experiments] Experiments section (and associated tables/figures): the reported gains in classification accuracy and post-intervention performance are not accompanied by capacity-matched ablations. If the latent concept graph module increases parameter count relative to the baseline CBM, the observed improvements could be explained by added expressivity rather than by correctly modeling concept correlations. A direct comparison (e.g., baseline CBM augmented with an equivalent number of unstructured parameters) is required to substantiate the central claim that the graph structure itself drives the benefits.
  2. [Method] §3 (Method): the construction of the latent concept graph is described at a high level, but the paper does not specify how the graph edges or adjacency matrix are learned or regularized, nor whether the graph parameters are frozen during intervention experiments. This detail is load-bearing for the intervention-effectiveness claim.
minor comments (2)
  1. [Abstract] Abstract: the phrase 'superior in image classification tasks' is vague; quantitative deltas, dataset names, and baseline comparisons should be stated explicitly.
  2. Notation: the distinction between the original CBM concept vector and the graph-augmented representation is not always clear in equations; consistent symbols would improve readability.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive and detailed feedback. We address each major comment point by point below, agreeing where additional controls and clarifications are warranted, and outlining the revisions we will make.

read point-by-point responses

  1. Referee: [Experiments] Experiments section (and associated tables/figures): the reported gains in classification accuracy and post-intervention performance are not accompanied by capacity-matched ablations. If the latent concept graph module increases parameter count relative to the baseline CBM, the observed improvements could be explained by added expressivity rather than by correctly modeling concept correlations. A direct comparison (e.g., baseline CBM augmented with an equivalent number of unstructured parameters) is required to substantiate the central claim that the graph structure itself drives the benefits.

    Authors: We agree that capacity-matched ablations are necessary to isolate the contribution of the graph structure from potential gains due to increased model capacity. In the revised manuscript we will add direct comparisons in which the baseline CBM is augmented with an equivalent number of unstructured parameters (for example by expanding hidden-layer dimensions or inserting additional fully-connected layers) to match the parameter count of the latent concept graph module. Updated tables and figures will report these results alongside the original experiments. revision: yes

  2. Referee: [Method] §3 (Method): the construction of the latent concept graph is described at a high level, but the paper does not specify how the graph edges or adjacency matrix are learned or regularized, nor whether the graph parameters are frozen during intervention experiments. This detail is load-bearing for the intervention-effectiveness claim.

    Authors: We acknowledge that the current description in §3 is high-level and omits explicit details on how the adjacency matrix and edges are learned or regularized, as well as the treatment of graph parameters during interventions. This information is important for reproducibility and for interpreting the intervention results. In the revised version we will expand §3 with a precise account of the graph-construction procedure, including the learning mechanism, any regularization terms, and whether graph parameters remain frozen during the intervention experiments. revision: yes

Circularity Check

0 steps flagged

No significant circularity in the proposed architectural extension or empirical claims

full rationale

The paper proposes GraphCBMs as an extension to standard Concept Bottleneck Models by adding latent concept graphs to capture correlations among concepts. This construction is motivated by a stated premise about intrinsic concept structure rather than derived from any equation or prior result within the paper. All claimed benefits—superior classification performance, more effective interventions, and robustness—are presented as outcomes of experiments on real-world image tasks, without any fitted parameters, predictions, or first-principles results that reduce to the inputs by construction. No self-citations, uniqueness theorems, or ansatzes are invoked in a load-bearing way in the provided text. The derivation chain is therefore self-contained as an empirical architectural proposal.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

The central proposal rests on the domain assumption that concepts possess an intrinsic correlated structure that can be usefully represented as a latent graph; no free parameters or new invented entities with independent evidence are specified in the abstract.

axioms (1)
  • domain assumption Concepts are generally correlated such that changing one inherently impacts related concepts.
    Invoked directly in the abstract to motivate the limitation of existing CBMs and the need for graph-based modeling.
invented entities (1)
  • Latent concept graphs no independent evidence
    purpose: To model and facilitate hidden relationships among concepts within the bottleneck.
    Introduced as the core new component that combines with CBMs; no falsifiable external evidence is provided in the abstract.

pith-pipeline@v0.9.0 · 5697 in / 1308 out tokens · 27583 ms · 2026-05-18T21:57:10.810186+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

  • IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear
    ?
    unclear

    Relation between the paper passage and the cited Recognition theorem.

    the set of concepts in CBMs often has an intrinsic structure where concepts are generally correlated: changing one concept will inherently impact its related concepts... constructing latent concept graphs... GNN Message Passing... V^l_emb = σ( D̃^{-1/2} Ã^l D̃^{-1/2} [V^{l-1}_act ⊙ V^{l-1}_emb] )

  • IndisputableMonolith/Foundation/RealityFromDistinction.lean reality_from_one_distinction unclear
    ?
    unclear

    Relation between the paper passage and the cited Recognition theorem.

    We hypothesize the existence of a unified, input-independent concept graph that encodes prior semantic knowledge... contrastive regularization term based on the normalized temperature-scaled cross-entropy loss

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

57 extracted references · 57 canonical work pages

  1. [1]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770–778, 2016

    work page 2016

  2. [2]

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

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021

    work page 2021

  3. [3]

    Mlp-mixer: An all-mlp architecture for vision

    Ilya O Tolstikhin, Neil Houlsby, Alexander Kolesnikov, Lucas Beyer, Xiaohua Zhai, Thomas Unterthiner, Jessica Yung, Andreas Steiner, Daniel Keysers, Jakob Uszkoreit, et al. Mlp-mixer: An all-mlp architecture for vision. Advances in neural information processing systems, 34:24261–24272, 2021

    work page 2021

  4. [4]

    Attention is all you need

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017

    work page 2017

  5. [5]

    BERT: Pre-training of deep bidirectional transformers for language understanding

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of deep bidirectional transformers for language understanding. In Jill Burstein, Christy Doran, and Thamar Solorio, editors,Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human 9 Language Technologies, Vo...

    work page 2019

  6. [6]

    Association for Computational Linguistics

    work page

  7. [7]

    Language models are few-shot learners

    Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Nee- lakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877–1901, 2020

    work page 1901

  8. [8]

    Exploring the limits of transfer learning with a unified text-to-text transformer

    Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of machine learning research, 21(140):1–67, 2020

    work page 2020

  9. [9]

    Kipf and Max Welling

    Thomas N. Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. In International Conference on Learning Representations, 2017

    work page 2017

  10. [10]

    Graph contrastive learning with augmentations

    Yuning You, Tianlong Chen, Yongduo Sui, Ting Chen, Zhangyang Wang, and Yang Shen. Graph contrastive learning with augmentations. Advances in neural information processing systems, 33:5812–5823, 2020

    work page 2020

  11. [11]

    Deep graph learning: Foundations, advances and applications

    Yu Rong, Tingyang Xu, Junzhou Huang, Wenbing Huang, Hong Cheng, Yao Ma, Yiqi Wang, Tyler Derr, Lingfei Wu, and Tengfei Ma. Deep graph learning: Foundations, advances and applications. In Proceedings of the 26th ACM SIGKDD international conference on knowledge discovery & data mining, pages 3555–3556, 2020

    work page 2020

  12. [12]

    Concept bottleneck models

    Pang Wei Koh, Thao Nguyen, Yew Siang Tang, Stephen Mussmann, Emma Pierson, Been Kim, and Percy Liang. Concept bottleneck models. In International conference on machine learning, pages 5338–5348. PMLR, 2020

    work page 2020

  13. [13]

    Neural representations for object perception: structure, category, and adaptive coding

    Zoe Kourtzi and Charles E Connor. Neural representations for object perception: structure, category, and adaptive coding. Annual review of neuroscience, 34(1):45–67, 2011

    work page 2011

  14. [14]

    Concept embedding models: Beyond the accuracy-explainability trade-off

    Mateo Espinosa Zarlenga, Pietro Barbiero, Gabriele Ciravegna, Giuseppe Marra, Francesco Giannini, Michelan- gelo Diligenti, Zohreh Shams, Frederic Precioso, Stefano Melacci, Adrian Weller, Pietro Lio, and Mateja Jamnik. Concept embedding models: Beyond the accuracy-explainability trade-off. Advances in Neural Information Processing Systems, 35, 2022

    work page 2022

  15. [15]

    Probabilistic concept bottleneck models

    Eunji Kim, Dahuin Jung, Sangha Park, Siwon Kim, and Sungroh Yoon. Probabilistic concept bottleneck models. International conference on machine learning, 2023

    work page 2023

  16. [16]

    Post-hoc concept bottleneck models

    Mert Yuksekgonul, Maggie Wang, and James Zou. Post-hoc concept bottleneck models. In The Eleventh International Conference on Learning Representations, 2023

    work page 2023

  17. [17]

    Visual objects in context

    Moshe Bar. Visual objects in context. Nature Reviews Neuroscience, 5(8):617–629, 2004

    work page 2004

  18. [18]

    The role of context in object recognition

    Aude Oliva and Antonio Torralba. The role of context in object recognition. Trends in cognitive sciences , 11(12):520–527, 2007

    work page 2007

  19. [19]

    Addressing leakage in concept bottleneck models

    Marton Havasi, Sonali Parbhoo, and Finale Doshi-Velez. Addressing leakage in concept bottleneck models. Advances in Neural Information Processing Systems, 35:23386–23397, 2022

    work page 2022

  20. [20]

    Energy-based concept bottleneck models: Unifying prediction, concept intervention, and probabilistic interpretations

    Xinyue Xu, Yi Qin, Lu Mi, Hao Wang, and Xiaomeng Li. Energy-based concept bottleneck models: Unifying prediction, concept intervention, and probabilistic interpretations. In The Twelfth International Conference on Learning Representations, 2024

    work page 2024

  21. [21]

    Nguyen, and Tsui-Wei Weng

    Tuomas Oikarinen, Subhro Das, Lam M. Nguyen, and Tsui-Wei Weng. Label-free concept bottleneck models. In The Eleventh International Conference on Learning Representations, 2023

    work page 2023

  22. [22]

    Language in a bottle: Language model guided concept bottlenecks for interpretable image classification

    Yue Yang, Artemis Panagopoulou, Shenghao Zhou, Daniel Jin, Chris Callison-Burch, and Mark Yatskar. Language in a bottle: Language model guided concept bottlenecks for interpretable image classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19187–19197, 2023

    work page 2023

  23. [23]

    Learning concise and descriptive attributes for visual recognition

    An Yan, Yu Wang, Yiwu Zhong, Chengyu Dong, Zexue He, Yujie Lu, William Yang Wang, Jingbo Shang, and Julian McAuley. Learning concise and descriptive attributes for visual recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 3090–3100, October 2023

    work page 2023

  24. [24]

    Learning graphs from data: A signal representation perspective

    Xiaowen Dong, Dorina Thanou, Michael Rabbat, and Pascal Frossard. Learning graphs from data: A signal representation perspective. IEEE Signal Processing Magazine, 36(3):44–63, 2019

    work page 2019

  25. [25]

    DAG-GNN: DAG structure learning with graph neural networks

    Yue Yu, Jie Chen, Tian Gao, and Mo Yu. DAG-GNN: DAG structure learning with graph neural networks. In Kamalika Chaudhuri and Ruslan Salakhutdinov, editors, Proceedings of the 36th International Conference on Machine Learning, volume 97 of Proceedings of Machine Learning Research, pages 7154–7163. PMLR, 09–15 Jun 2019

    work page 2019

  26. [26]

    Learning discrete structures for graph neural networks

    Luca Franceschi, Mathias Niepert, Massimiliano Pontil, and Xiao He. Learning discrete structures for graph neural networks. In International conference on machine learning, pages 1972–1982. PMLR, 2019. 10

    work page 1972

  27. [27]

    Neural relational inference for interacting systems

    Thomas Kipf, Ethan Fetaya, Kuan-Chieh Wang, Max Welling, and Richard Zemel. Neural relational inference for interacting systems. In International conference on machine learning, pages 2688–2697. PMLR, 2018

    work page 2018

  28. [28]

    Discrete graph structure learning for forecasting multiple time series

    Chao Shang, Jie Chen, and Jinbo Bi. Discrete graph structure learning for forecasting multiple time series. In International Conference on Learning Representations, 2020

    work page 2020

  29. [29]

    Differentiable graph module (dgm) for graph convolutional networks

    Anees Kazi, Luca Cosmo, Seyed-Ahmad Ahmadi, Nassir Navab, and Michael M Bronstein. Differentiable graph module (dgm) for graph convolutional networks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(2):1606–1617, 2022

    work page 2022

  30. [30]

    Federated learning of models pre-trained on different features with consensus graphs

    Tengfei Ma, Trong Nghia Hoang, and Jie Chen. Federated learning of models pre-trained on different features with consensus graphs. In Uncertainty in Artificial Intelligence, pages 1336–1346. PMLR, 2023

    work page 2023

  31. [31]

    Augmentations in hypergraph contrastive learning: Fabricated and generative

    Tianxin Wei, Yuning You, Tianlong Chen, Yang Shen, Jingrui He, and Zhangyang Wang. Augmentations in hypergraph contrastive learning: Fabricated and generative. In S. Koyejo, S. Mohamed, A. Agarwal, D. Belgrave, K. Cho, and A. Oh, editors, Advances in Neural Information Processing Systems, volume 35, pages 1909–1922. Curran Associates, Inc., 2022

    work page 1909

  32. [32]

    A simple framework for contrastive learning of visual representations

    Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597–1607. PMLR, 2020

    work page 2020

  33. [33]

    struc2vec: Learning node representations from structural identity

    Leonardo FR Ribeiro, Pedro HP Saverese, and Daniel R Figueiredo. struc2vec: Learning node representations from structural identity. In Proceedings of the 23rd ACM SIGKDD international conference on knowledge discovery and data mining, pages 385–394, 2017

    work page 2017

  34. [34]

    Infograph: Unsupervised and semi-supervised graph-level representation learning via mutual information maximization

    Fan-Yun Sun, Jordan Hoffman, Vikas Verma, and Jian Tang. Infograph: Unsupervised and semi-supervised graph-level representation learning via mutual information maximization. InInternational Conference on Learning Representations, 2019

    work page 2019

  35. [35]

    Graph contrastive learning with augmentations

    Yuning You, Tianlong Chen, Yongduo Sui, Ting Chen, Zhangyang Wang, and Yang Shen. Graph contrastive learning with augmentations. In H. Larochelle, M. Ranzato, R. Hadsell, M. F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems, volume 33, pages 5812–5823. Curran Associates, Inc., 2020

    work page 2020

  36. [36]

    Multi-level contrastive learning framework for sequential recommendation

    Ziyang Wang, Huoyu Liu, Wei Wei, Yue Hu, Xian-Ling Mao, Shaojian He, Rui Fang, and Dangyang Chen. Multi-level contrastive learning framework for sequential recommendation. In Proceedings of the 31st ACM International Conference on Information & Knowledge Management, CIKM ’22, page 2098–2107, New York, NY , USA, 2022. Association for Computing Machinery

    work page 2098

  37. [37]

    Understanding contrastive representation learning through alignment and uniformity on the hypersphere

    Tongzhou Wang and Phillip Isola. Understanding contrastive representation learning through alignment and uniformity on the hypersphere. In International conference on machine learning, pages 9929–9939. PMLR, 2020

    work page 2020

  38. [38]

    C. Wah, S. Branson, P. Welinder, P. Perona, and S. Belongie. The caltech-ucsd birds-200-2011 dataset. Technical Report CNS-TR-2011-001, California Institute of Technology, 2011

    work page 2011

  39. [39]

    Automated flower classification over a large number of classes

    Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In Indian Conference on Computer Vision, Graphics and Image Processing, Dec 2008

    work page 2008

  40. [40]

    Zero-shot learning—a comprehensive evaluation of the good, the bad and the ugly

    Yongqin Xian, Christoph H Lampert, Bernt Schiele, and Zeynep Akata. Zero-shot learning—a comprehensive evaluation of the good, the bad and the ugly. IEEE transactions on pattern analysis and machine intelligence, 41(9):2251–2265, 2018

    work page 2018

  41. [41]

    Learning multiple layers of features from tiny images

    Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009

    work page 2009

  42. [42]

    The HAM10000 dataset, a large collection of multi-source dermatoscopic images of common pigmented skin lesions, 2018

    Philipp Tschandl. The HAM10000 dataset, a large collection of multi-source dermatoscopic images of common pigmented skin lesions, 2018

    work page 2018

  43. [43]

    Chexpert: A large chest radiograph dataset with uncertainty labels and expert comparison

    Jeremy Irvin, Pranav Rajpurkar, Michael Ko, Yifan Yu, Silviana Ciurea-Ilcus, Chris Chute, Henrik Marklund, Behzad Haghgoo, Robyn Ball, Katie Shpanskaya, et al. Chexpert: A large chest radiograph dataset with uncertainty labels and expert comparison. In Proceedings of the AAAI conference on artificial intelligence, volume 33, pages 590–597, 2019

    work page 2019

  44. [44]

    Learning transferable visual models from natural language supervision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In Marina Meila and Tong Zhang, editors, Proceedings of the 38th International Conference on Machine...

    work page 2021

  45. [45]

    Learning to exploit temporal structure for biomedical vision-language processing

    Shruthi Bannur, Stephanie Hyland, Qianchu Liu, Fernando Perez-Garcia, Maximilian Ilse, Daniel C Castro, Benedikt Boecking, Harshita Sharma, Kenza Bouzid, Anja Thieme, et al. Learning to exploit temporal structure for biomedical vision-language processing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15016–150...

    work page 2023

  46. [46]

    Learning bottleneck concepts in image classification

    Bowen Wang, Liangzhi Li, Yuta Nakashima, and Hajime Nagahara. Learning bottleneck concepts in image classification. In Proceedings of the ieee/cvf conference on computer vision and pattern recognition , pages 10962–10971, 2023

    work page 2023

  47. [47]

    Panousis, Dino Ienco, and Diego Marcos

    Konstantinos P. Panousis, Dino Ienco, and Diego Marcos. Sparse linear concept discovery models. In 2023 IEEE/CVF International Conference on Computer Vision Workshops (ICCVW), pages 2759–2763, 2023

    work page 2023

  48. [48]

    Incremental residual concept bottleneck models

    Chenming Shang, Shiji Zhou, Hengyuan Zhang, Xinzhe Ni, Yujiu Yang, and Yuwang Wang. Incremental residual concept bottleneck models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11030–11040, June 2024

    work page 2024

  49. [49]

    Stochastic concept bottleneck models

    Moritz Vandenhirtz, Sonia Laguna, Riˇcards Marcinkeviˇcs, and Julia E V ogt. Stochastic concept bottleneck models. In Thirty-eighth Conference on Neural Information Processing Systems, 2024

    work page 2024

  50. [50]

    A closer look at the intervention procedure of concept bottleneck models

    Sungbin Shin, Yohan Jo, Sungsoo Ahn, and Namhoon Lee. A closer look at the intervention procedure of concept bottleneck models. In Andreas Krause, Emma Brunskill, Kyunghyun Cho, Barbara Engelhardt, Sivan Sabato, and Jonathan Scarlett, editors, Proceedings of the 40th International Conference on Machine Learning, volume 202 of Proceedings of Machine Learni...

    work page 2023

  51. [51]

    Training language models to follow instructions with human feedback

    Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, John Schulman, Jacob Hilton, Fraser Kelton, Luke Miller, Maddie Simens, Amanda Askell, Peter Welinder, Paul F Christiano, Jan Leike, and Ryan Lowe. Training language models to follow instructions with human feedbac...

    work page 2022

  52. [52]

    Linear explanations for individual neurons

    Tuomas Oikarinen and Tsui-Wei Weng. Linear explanations for individual neurons. In Ruslan Salakhutdinov, Zico Kolter, Katherine Heller, Adrian Weller, Nuria Oliver, Jonathan Scarlett, and Felix Berkenkamp, editors, Proceedings of the 41st International Conference on Machine Learning, volume 235 of Proceedings of Machine Learning Research, pages 38639–3866...

    work page 2024

  53. [53]

    Torchvision: Pytorch’s computer vision library.https://github

    TorchVision maintainers and contributors. Torchvision: Pytorch’s computer vision library.https://github. com/pytorch/vision, 2016

    work page 2016

  54. [54]

    Semantic-aware scene recognition

    Alejandro López-Cifuentes, Marcos Escudero-Viñolo, Jesús Bescós, and Álvaro García-Martín. Semantic-aware scene recognition. Pattern Recognition, page 107256, 2020

    work page 2020

  55. [55]

    No Finding

    Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248–255, 2009. A Dataset Introduction • Caltech-UCSD Birds-200-2011, CUB [ 37]: CUB is the most widely-used dataset for fine-grained visual categorization task...

    work page 2009

  56. [56]

    Pleural Other

    Heart failure can lead to pleural effusion (fluid in the pleural space), which may manifest as a "Pleural Other" abnormality. 18

    work page

  57. [57]

    Aortic aneurysm or dissection may affect surrounding pleural structures due to proximity, causing pleural thickening or effusions. Malignancies: Tumors in the mediastinum (e.g., lymphomas or metastatic disease) can enlarge the cardiomediastinum and simultaneously invade or affect the pleura, leading to pleural abnormalities. Infections and inflammatory co...

    work page