pith. sign in

arxiv: 2606.28112 · v1 · pith:LAGQRCMAnew · submitted 2026-06-26 · 💻 cs.CV · cs.AI

BiDeMem: Bidirectional Degradation Memory for Explainable Image Restoration

Xinrui Wu , Lichen Huang This is my paper

Pith reviewed 2026-06-29 03:55 UTC · model grok-4.3

classification 💻 cs.CV cs.AI
keywords image restorationdegradation memoryexplainable AImemory retrievalbidirectional modelmulti-degradationprior interventionNAFNet
0
0 comments X

The pith

Bidirectional memory retrieves degradation slots to support both image restoration and falsifiable explanations of degradation type.

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

Degradation conditions in image restoration are usually judged only by final PSNR, a weak test that cannot separate true semantic priors from added capacity or dataset shortcuts. BiDeMem builds a memory whose query, formed from restoration features and input statistics, selects a top-k set of slots. Those slot identities feed the restoration path at inference time and a separate forward-degradation explanation path available only in training. In a controlled multi-degradation NAFNet setup the complete bidirectional model exceeds correction-head, dense-query-prior, and static-global-prior baselines by 0.2588 dB, 0.2586 dB, and 0.2839 dB. Intervention probes that replace correct slots with wrong-prior or native-prior slots keep restoration quality intact while raising sensitivity to the identity of the retrieved slots.

Core claim

BiDeMem retrieves a compact top-k subset of memory slots from a query built on restoration features and input statistics; the same slot identities drive the restoration path at inference and a training-only forward-degradation explanation path. Intervention probes confirm that this arrangement preserves restoration quality while increasing measurable sensitivity to wrong-prior and native-prior swaps, outperforming correction-head, dense-query-prior, and static-global-prior variants by 0.2588, 0.2586, and 0.2839 dB respectively.

What carries the argument

Bidirectional degradation memory: query-retrieved top-k slots whose identities serve both a restoration path and a training-only explanation path.

If this is right

  • The full bidirectional model outperforms a correction-head variant by 0.2588 dB.
  • It outperforms a dense-query-prior variant by 0.2586 dB.
  • It outperforms a static-global-prior variant by 0.2839 dB.
  • Strong residual supervision and a wider degradation head both fall short of the bidirectional memory model.
  • Intervention on slot identity raises wrong-prior and native-prior sensitivity while leaving restoration quality intact.

Where Pith is reading between the lines

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

  • If slot identities reliably encode degradation type, the same memory could be used after training to diagnose which degradation a model is mishandling on a given input.
  • The training-only explanation path offers a template for adding interpretability modules that impose no cost at inference time.
  • The retrieval-plus-intervention design could be tested on other conditional vision tasks where the semantic content of a condition must be verified independently of final task accuracy.

Load-bearing premise

That the identity of a retrieved memory slot can be treated as a semantically meaningful representation of degradation type so that swapping it produces interpretable, measurable changes rather than mere capacity effects.

What would settle it

An experiment that forces the model to use a wrong-prior slot on images whose degradation type is known in advance and checks whether PSNR or output statistics remain statistically unchanged from the correct-prior case.

Figures

Figures reproduced from arXiv: 2606.28112 by Lichen Huang, Xinrui Wu.

Figure 1
Figure 1. Figure 1: Overview of BiDeMem. A degraded image is encoded by a restoration backbone. Evidence from the bottleneck feature and input statistics forms a query that retrieves a compact top-k subset from a degradation memory. The same selected slot identity conditions the restoration decoder during inference and, during training only, drives a forward degradation explanation branch. The figure emphasizes the paper’s ce… view at source ↗
Figure 2
Figure 2. Figure 2: Qualitative restoration examples. Each row compares the low-quality input, ground truth, NAFNet, NAFNet-wide, Rank Memory, and BiRank Memory. Yellow boxes highlight local regions where the memory variants better recover contrast and structure. The numbers in the headers are per-example PSNR. tors, but their priors are usually tied to a specific degradation or evaluated mainly through endpoint quality. Gene… view at source ↗
Figure 3
Figure 3. Figure 3: Intervention evidence. The bars summarize counterfac￾tual prior tests. Larger gaps or drops mean that the output depends more strongly on the correct retrieved prior. BiRank Memory especially increases wrong-prior sensitivity and the native/non￾native prior gap while preserving restoration accuracy [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
read the original abstract

Degradation-aware prompts, conditions, and latent priors are increasingly used in image restoration, yet they are usually judged by a single endpoint: whether the restored image obtains higher PSNR. This is a weak test of semantics. A condition can help by adding capacity, acting as a global correction bias, or exploiting dataset shortcuts, without becoming an interpretable degradation prior. We propose BiDeMem, a bidirectional degradation memory for explainable image restoration. A query built from restoration features and input statistics retrieves a compact top-k subset of memory slots. The same selected slot identity supports the restoration path at inference time and a training-only forward-degradation explanation path. The study centers on verifiability in a controlled multi-degradation NAFNet setting. New controls separate the gain from a correction head alone, a dense query prior, and a static global prior: these variants are 0.2588, 0.2586, and 0.2839 dB below BiRank, respectively. Strong residual supervision and a wider degradation head also remain below the full bidirectional memory model. Intervention probes show that BiRank preserves restoration quality while increasing wrong-prior and native-prior sensitivity, framing degradation memory as both a restoration module and a falsifiable explanation mechanism.

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

3 major / 1 minor

Summary. The paper proposes BiDeMem, a bidirectional degradation memory for explainable image restoration. A query derived from restoration features and input statistics retrieves a top-k subset of memory slots; the slot identities are used both in the restoration forward pass and in a training-only forward-degradation explanation path. Experiments are conducted in a controlled multi-degradation NAFNet setting. The full bidirectional model is reported to outperform three explicit variant baselines (correction head, dense query prior, static global prior) by 0.2588 dB, 0.2586 dB, and 0.2839 dB respectively. Intervention probes (wrong-prior and native-prior swaps) are claimed to increase sensitivity while preserving restoration quality, positioning the memory as both a performance module and a falsifiable explanation mechanism.

Significance. If the central claim holds—that retrieved memory slot identities constitute a semantically meaningful and falsifiable representation of degradation type rather than an additional learned bias—the work would provide a concrete methodological advance in evaluating degradation-aware priors beyond endpoint PSNR. The explicit construction of multiple ablation variants and the use of intervention probes are positive design choices that ground the comparison. The small reported margins and absence of supporting validation details, however, limit the immediate impact.

major comments (3)
  1. [Abstract] Abstract: the reported PSNR gaps (0.2588 dB, 0.2586 dB, 0.2839 dB) are presented without error bars, standard deviations, or statistical significance tests across multiple runs or dataset splits; this directly affects the claim that the bidirectional model is superior to the three named variants.
  2. [Abstract] Abstract (and method description): no specification is given for memory-slot initialization, population rules, update mechanism, or the precise contents of each slot; without these details the assertion that slot identities function as a falsifiable degradation representation cannot be evaluated.
  3. [Abstract] Abstract (intervention probes paragraph): the claim that BiRank 'increases wrong-prior and native-prior sensitivity' while preserving quality is consistent with the prior acting simply as an extra learned input channel; the manuscript supplies no auxiliary check (e.g., clustering of slot identities against known degradation categories or alignment metrics) that would distinguish semantic representation from capacity augmentation.
minor comments (1)
  1. [Abstract] The term 'BiRank' appears in the abstract without prior definition; clarify its relation to BiDeMem.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the thoughtful and constructive review. The comments highlight important aspects of statistical rigor, methodological transparency, and validation of semantic claims. We address each major comment below and commit to revisions that strengthen the manuscript without altering its core contributions.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the reported PSNR gaps (0.2588 dB, 0.2586 dB, 0.2839 dB) are presented without error bars, standard deviations, or statistical significance tests across multiple runs or dataset splits; this directly affects the claim that the bidirectional model is superior to the three named variants.

    Authors: We agree that reporting variability and significance is essential for robust claims. The current results reflect single-run evaluations in the controlled NAFNet setting. In the revision we will rerun all experiments over five independent random seeds, report mean PSNR with standard deviations, and include paired t-test p-values comparing BiDeMem against each baseline. This will directly address the concern about the small margins. revision: yes

  2. Referee: [Abstract] Abstract (and method description): no specification is given for memory-slot initialization, population rules, update mechanism, or the precise contents of each slot; without these details the assertion that slot identities function as a falsifiable degradation representation cannot be evaluated.

    Authors: The method section describes a fixed-size learnable memory bank whose slots are retrieved by top-k similarity and whose identities are used in both paths, but we concede that explicit initialization, population, and update details are insufficiently specified. We will add a dedicated subsection detailing: random normal initialization scaled by slot dimension, fixed population of 128 slots, end-to-end gradient updates with no explicit population rule beyond back-propagation, and slot contents as 256-dimensional embedding vectors. These additions will make the falsifiability claim evaluable. revision: yes

  3. Referee: [Abstract] Abstract (intervention probes paragraph): the claim that BiRank 'increases wrong-prior and native-prior sensitivity' while preserving quality is consistent with the prior acting simply as an extra learned input channel; the manuscript supplies no auxiliary check (e.g., clustering of slot identities against known degradation categories or alignment metrics) that would distinguish semantic representation from capacity augmentation.

    Authors: The intervention probes (wrong-prior and native-prior swaps) already show that swapping slot identities alters restoration behavior in a degradation-dependent way while quality remains comparable, which goes beyond generic capacity. Nevertheless, we accept that auxiliary checks would better separate semantic representation from bias. In revision we will add (i) k-means clustering of learned slot identities colored by ground-truth degradation category and (ii) normalized mutual information between slot assignments and degradation labels, providing quantitative evidence that the memory captures more than an extra channel. revision: yes

Circularity Check

0 steps flagged

No significant circularity in derivation chain

full rationale

The paper's central claims rest on empirical comparisons to explicitly constructed variant architectures (correction head, dense query prior, static global prior) and intervention probes within a controlled NAFNet multi-degradation setting. These provide independent external benchmarks rather than reducing to self-definition or fitted inputs renamed as predictions. No equations or self-citations are shown that force the explanatory status of memory slots by construction; the bidirectional path is tested for sensitivity while preserving restoration quality, keeping the argument self-contained against those controls.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 1 invented entities

The central claim rests on the introduction of a new memory structure whose slots are assumed to encode degradation semantics; no external benchmarks or machine-checked proofs are mentioned.

free parameters (1)
  • top-k
    Number of memory slots retrieved by the query; chosen as a hyper-parameter to balance capacity and compactness.
axioms (1)
  • domain assumption Memory slots populated from restoration features and input statistics can represent distinct degradation types
    Invoked when the query selects slots that are later used for both restoration and degradation explanation.
invented entities (1)
  • Bidirectional degradation memory slots no independent evidence
    purpose: Compact storage of degradation information that supports both restoration conditioning and a training-only explanation path
    New postulated structure introduced by the model; no independent evidence outside the paper is provided.

pith-pipeline@v0.9.1-grok · 5750 in / 1628 out tokens · 61098 ms · 2026-06-29T03:55:14.667410+00:00 · methodology

discussion (0)

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

Reference graph

Works this paper leans on

47 extracted references · 4 canonical work pages · 4 internal anchors

  1. [1]

    Beyond a gaussian denoiser: Residual learning of deep cnn for image denoising.IEEE transactions on image processing, 26(7):3142–3155, 2017

    Kai Zhang, Wangmeng Zuo, Yunjin Chen, Deyu Meng, and Lei Zhang. Beyond a gaussian denoiser: Residual learning of deep cnn for image denoising.IEEE transactions on image processing, 26(7):3142–3155, 2017. 1, 2

    2017

  2. [2]

    Benchmarking single- image dehazing and beyond.IEEE transactions on image processing, 28(1):492–505, 2018

    Boyi Li, Wenqi Ren, Dengpan Fu, Dacheng Tao, Dan Feng, Wenjun Zeng, and Zhangyang Wang. Benchmarking single- image dehazing and beyond.IEEE transactions on image processing, 28(1):492–505, 2018. 2, 5

    2018

  3. [3]

    Deep joint rain detection and removal from a single image

    Wenhan Yang, Robby T Tan, Jiashi Feng, Jiaying Liu, Zong- ming Guo, and Shuicheng Yan. Deep joint rain detection and removal from a single image. InProceedings of the IEEE conference on computer vision and pattern recogni- tion, pages 1357–1366, 2017. 2, 5

    2017

  4. [4]

    Multi-stage progressive image restoration

    Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, Ming-Hsuan Yang, and Ling Shao. Multi-stage progressive image restoration. InPro- ceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14821–14831, 2021. 3

    2021

  5. [5]

    Restormer: Efficient transformer for high-resolution image restoration

    Syed Waqas Zamir, Aditya Arora, Salman Khan, Mu- nawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5728–5739,

  6. [6]

    Simple baselines for image restoration

    Liangyu Chen, Xiaojie Chu, Xiangyu Zhang, and Jian Sun. Simple baselines for image restoration. InEuropean confer- ence on computer vision, pages 17–33. Springer, 2022. 1, 3, 5

    2022

  7. [7]

    All-in-one image restoration for unknown corruption

    Boyun Li, Xiao Liu, Peng Hu, Zhongqin Wu, Jiancheng Lv, and Xi Peng. All-in-one image restoration for unknown corruption. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 17452– 17462, 2022. 1, 3, 7

    2022

  8. [8]

    Transweather: Transformer-based restoration of im- ages degraded by adverse weather conditions

    Jeya Maria Jose Valanarasu, Rajeev Yasarla, and Vishal M Patel. Transweather: Transformer-based restoration of im- ages degraded by adverse weather conditions. InProceed- ings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2353–2363, 2022. 3

    2022

  9. [9]

    Promptir: Prompting for all-in- one image restoration.Advances in neural information pro- cessing systems, 36:71275–71293, 2023

    Vaishnav Potlapalli, Syed Waqas Zamir, Salman H Khan, and Fahad Shahbaz Khan. Promptir: Prompting for all-in- one image restoration.Advances in neural information pro- cessing systems, 36:71275–71293, 2023. 1, 3, 4, 7

    2023

  10. [10]

    Ffdnet: Toward a fast and flexible solution for cnn-based image denoising

    Kai Zhang, Wangmeng Zuo, and Lei Zhang. Ffdnet: Toward a fast and flexible solution for cnn-based image denoising. IEEE Transactions on Image Processing, 27(9):4608–4622,

  11. [11]

    Toward convolutional blind denoising of real pho- tographs

    Shi Guo, Zifei Yan, Kai Zhang, Wangmeng Zuo, and Lei Zhang. Toward convolutional blind denoising of real pho- tographs. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1712–1722,

  12. [12]

    Noise2Noise: Learning Image Restoration without Clean Data

    Jaakko Lehtinen, Jacob Munkberg, Jon Hasselgren, Samuli Laine, Tero Karras, Miika Aittala, and Timo Aila. Noise2noise: Learning image restoration without clean data. arXiv preprint arXiv:1803.04189, 2018. 2

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  13. [13]

    Noise2void-learning denoising from single noisy images

    Alexander Krull, Tim-Oliver Buchholz, and Florian Jug. Noise2void-learning denoising from single noisy images. In Proceedings of the IEEE/CVF conference on computer vi- sion and pattern recognition, pages 2129–2137, 2019. 2

    2019

  14. [14]

    Single image haze removal using dark channel prior.IEEE transactions on pat- tern analysis and machine intelligence, 33(12):2341–2353,

    Kaiming He, Jian Sun, and Xiaoou Tang. Single image haze removal using dark channel prior.IEEE transactions on pat- tern analysis and machine intelligence, 33(12):2341–2353,

  15. [15]

    Dehazenet: An end-to-end system for single image haze removal.IEEE transactions on image process- ing, 25(11):5187–5198, 2016

    Bolun Cai, Xiangmin Xu, Kui Jia, Chunmei Qing, and Dacheng Tao. Dehazenet: An end-to-end system for single image haze removal.IEEE transactions on image process- ing, 25(11):5187–5198, 2016. 2

    2016

  16. [16]

    Aod-net: All-in-one dehazing network

    Boyi Li, Xiulian Peng, Zhangyang Wang, Jizheng Xu, and Dan Feng. Aod-net: All-in-one dehazing network. InPro- ceedings of the IEEE international conference on computer vision, pages 4770–4778, 2017. 2

    2017

  17. [17]

    Density-aware single image de-raining using a multi-stream dense network

    He Zhang and Vishal M Patel. Density-aware single image de-raining using a multi-stream dense network. InProceed- ings of the IEEE conference on computer vision and pattern recognition, pages 695–704, 2018. 2

    2018

  18. [18]

    Removing rain from single images via a deep detail network

    Xueyang Fu, Jiabin Huang, Delu Zeng, Yue Huang, Xinghao Ding, and John Paisley. Removing rain from single images via a deep detail network. InProceedings of the IEEE con- ference on computer vision and pattern recognition, pages 3855–3863, 2017. 2

    2017

  19. [19]

    Recurrent squeeze-and-excitation context aggregation net for single image deraining

    Xia Li, Jianlong Wu, Zhouchen Lin, Hong Liu, and Hongbin Zha. Recurrent squeeze-and-excitation context aggregation net for single image deraining. InProceedings of the Eu- ropean conference on computer vision (ECCV), pages 254– 269, 2018. 2

    2018

  20. [20]

    Spatial attentive single-image deraining with a high quality real rain dataset

    Tianyu Wang, Xin Yang, Ke Xu, Shaozhe Chen, Qiang Zhang, and Rynson WH Lau. Spatial attentive single-image deraining with a high quality real rain dataset. InProceed- ings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12270–12279, 2019. 2

    2019

  21. [21]

    Deblurgan: Blind mo- tion deblurring using conditional adversarial networks

    Orest Kupyn, V olodymyr Budzan, Mykola Mykhailych, Dmytro Mishkin, and Ji ˇr´ı Matas. Deblurgan: Blind mo- tion deblurring using conditional adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8183–8192, 2018. 2 8

    2018

  22. [22]

    Deblurgan-v2: Deblurring (orders-of-magnitude) faster and better

    Orest Kupyn, Tetiana Martyniuk, Junru Wu, and Zhangyang Wang. Deblurgan-v2: Deblurring (orders-of-magnitude) faster and better. InProceedings of the IEEE/CVF inter- national conference on computer vision, pages 8878–8887,

  23. [23]

    Enhanced deep residual networks for single image super-resolution

    Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Kyoung Mu Lee. Enhanced deep residual networks for single image super-resolution. InProceedings of the IEEE confer- ence on computer vision and pattern recognition workshops, pages 136–144, 2017. 2

    2017

  24. [24]

    Image super-resolution using very deep residual channel attention networks

    Yulun Zhang, Kunpeng Li, Kai Li, Lichen Wang, Bineng Zhong, and Yun Fu. Image super-resolution using very deep residual channel attention networks. InProceedings of the European conference on computer vision (ECCV), pages 286–301, 2018. 2

    2018

  25. [25]

    Learning enriched features for real image restoration and enhancement

    Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, Ming-Hsuan Yang, and Ling Shao. Learning enriched features for real image restoration and enhancement. InEuropean conference on computer vi- sion, pages 492–511. Springer, 2020. 3

    2020

  26. [26]

    Swinir: Image restoration us- ing swin transformer

    Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration us- ing swin transformer. InProceedings of the IEEE/CVF inter- national conference on computer vision, pages 1833–1844,

  27. [27]

    Uformer: A general u-shaped transformer for image restoration

    Zhendong Wang, Xiaodong Cun, Jianmin Bao, Wengang Zhou, Jianzhuang Liu, and Houqiang Li. Uformer: A general u-shaped transformer for image restoration. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 17683–17693, 2022. 3

    2022

  28. [28]

    Learning to see in the dark

    Chen Chen, Qifeng Chen, Jia Xu, and Vladlen Koltun. Learning to see in the dark. InProceedings of the IEEE con- ference on computer vision and pattern recognition, pages 3291–3300, 2018. 3

    2018

  29. [29]

    Zero-reference deep curve estimation for low-light image enhancement

    Chunle Guo, Chongyi Li, Jichang Guo, Chen Change Loy, Junhui Hou, Sam Kwong, and Runmin Cong. Zero-reference deep curve estimation for low-light image enhancement. In Proceedings of the IEEE/CVF conference on computer vi- sion and pattern recognition, pages 1780–1789, 2020. 3

    2020

  30. [30]

    Film: Visual reasoning with a general conditioning layer

    Ethan Perez, Florian Strub, Harm De Vries, Vincent Du- moulin, and Aaron Courville. Film: Visual reasoning with a general conditioning layer. InProceedings of the AAAI con- ference on artificial intelligence, 2018. 3

    2018

  31. [31]

    Mod- ulating early visual processing by language.Advances in neural information processing systems, 30, 2017

    Harm De Vries, Florian Strub, J ´er´emie Mary, Hugo Larochelle, Olivier Pietquin, and Aaron C Courville. Mod- ulating early visual processing by language.Advances in neural information processing systems, 30, 2017. 4

    2017

  32. [32]

    Dynamic filter networks.Advances in neural informa- tion processing systems, 29, 2016

    Xu Jia, Bert De Brabandere, Tinne Tuytelaars, and Luc V Gool. Dynamic filter networks.Advances in neural informa- tion processing systems, 29, 2016. 4

    2016

  33. [33]

    Memory Networks

    Jason Weston, Sumit Chopra, and Antoine Bordes. Memory networks.arXiv preprint arXiv:1410.3916, 2014. 4

    work page internal anchor Pith review Pith/arXiv arXiv 2014

  34. [34]

    End- to-end memory networks.Advances in neural information processing systems, 28, 2015

    Sainbayar Sukhbaatar, Jason Weston, Rob Fergus, et al. End- to-end memory networks.Advances in neural information processing systems, 28, 2015. 4

    2015

  35. [35]

    Neural Turing Machines

    Alex Graves, Greg Wayne, and Ivo Danihelka. Neural turing machines.arXiv preprint arXiv:1410.5401, 2014. 4

    work page internal anchor Pith review Pith/arXiv arXiv 2014

  36. [36]

    Hybrid computing using a neural network with dynamic external memory.Nature, 538(7626): 471–476, 2016

    Alex Graves, Greg Wayne, Malcolm Reynolds, Tim Harley, Ivo Danihelka, Agnieszka Grabska-Barwi ´nska, Ser- gio G ´omez Colmenarejo, Edward Grefenstette, Tiago Ra- malho, John Agapiou, et al. Hybrid computing using a neural network with dynamic external memory.Nature, 538(7626): 471–476, 2016. 4

    2016

  37. [37]

    Object- centric learning with slot attention.Advances in neural in- formation processing systems, 33:11525–11538, 2020

    Francesco Locatello, Dirk Weissenborn, Thomas Un- terthiner, Aravindh Mahendran, Georg Heigold, Jakob Uszkoreit, Alexey Dosovitskiy, and Thomas Kipf. Object- centric learning with slot attention.Advances in neural in- formation processing systems, 33:11525–11538, 2020. 4

    2020

  38. [38]

    Large memory layers with product keys.Advances in Neural Information Processing Systems, 32, 2019

    Guillaume Lample, Alexandre Sablayrolles, Marc’Aurelio Ranzato, Ludovic Denoyer, and Herv´e J´egou. Large memory layers with product keys.Advances in Neural Information Processing Systems, 32, 2019. 4

    2019

  39. [39]

    Grad-cam: visual explanations from deep networks via gradient-based localization.International journal of com- puter vision, 128(2):336–359, 2020

    Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Ba- tra. Grad-cam: visual explanations from deep networks via gradient-based localization.International journal of com- puter vision, 128(2):336–359, 2020. 4

    2020

  40. [40]

    Axiomatic attribution for deep networks

    Mukund Sundararajan, Ankur Taly, and Qiqi Yan. Axiomatic attribution for deep networks. InInternational conference on machine learning, pages 3319–3328. PMLR, 2017. 4

    2017

  41. [41]

    Interpretability be- yond feature attribution: Quantitative testing with concept activation vectors (tcav)

    Been Kim, Martin Wattenberg, Justin Gilmer, Carrie Cai, James Wexler, Fernanda Viegas, et al. Interpretability be- yond feature attribution: Quantitative testing with concept activation vectors (tcav). InInternational conference on ma- chine learning, pages 2668–2677. PMLR, 2018. 4

    2018

  42. [42]

    ” why should i trust you?” explaining the predictions of any classifier

    Marco Tulio Ribeiro, Sameer Singh, and Carlos Guestrin. ” why should i trust you?” explaining the predictions of any classifier. InProceedings of the 22nd ACM SIGKDD interna- tional conference on knowledge discovery and data mining, pages 1135–1144, 2016. 4

    2016

  43. [43]

    A unified approach to interpreting model predictions.Advances in neural informa- tion processing systems, 30, 2017

    Scott M Lundberg and Su-In Lee. A unified approach to interpreting model predictions.Advances in neural informa- tion processing systems, 30, 2017. 4

    2017

  44. [44]

    Counterfactual Explanations without Opening the Black Box: Automated Decisions and the GDPR

    Sandra Wachter, Brent Mittelstadt, and Chris Russell. Counterfactual explanations without opening the black box: Automated decisions and the gdpr.arXiv preprint arXiv:1711.00399, 2017. 4

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  45. [45]

    Explanations based on the missing: Towards contrastive ex- planations with pertinent negatives.Advances in neural in- formation processing systems, 31, 2018

    Amit Dhurandhar, Pin-Yu Chen, Ronny Luss, Chun-Chen Tu, Paishun Ting, Karthikeyan Shanmugam, and Payel Das. Explanations based on the missing: Towards contrastive ex- planations with pertinent negatives.Advances in neural in- formation processing systems, 31, 2018. 4

    2018

  46. [46]

    A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics

    David Martin, Charless Fowlkes, Doron Tal, and Jitendra Malik. A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics. InProceedings of the IEEE International Conference on Computer Vision, pages 416– 423, 2001. 5

    2001

  47. [47]

    Sin- gle image super-resolution from transformed self-exemplars

    Jia-Bin Huang, Abhishek Singh, and Narendra Ahuja. Sin- gle image super-resolution from transformed self-exemplars. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5197–5206, 2015. 5 9

    2015