REVIEW 5 major objections 5 minor 33 references
Fill in the blanks: Rethinking Interpretability in vision
T0 review · 5 major / 5 minor · reviewed 2026-08-12 · deepseek-v4-flash
Pith's one-line read By masking part of an image and asking a trained classifier to fill in the missing pixels via a gradient update, this paper claims to expose the prototypical structure the model has learned for each class, without training any extra…
desk verdict A plausible mask-filling interpretability trick that is currently under-evidenced because its only quantitative test is circular. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The central object is the gradient update vector $\Delta x = -\gamma \nabla_x L(f_\theta(M(x)), y)$, computed by backpropagating the classification loss with respect to the pixels of a masked image. This vector is interpreted as the model's suggestion for how to change the input so the masked image better matches the predicted class. A thresholding transform $T$ retains only updates inside the masked regions and binarizes them for contrast, yielding the visual fill-in $M(x)+T(\Delta x)$. The argument that this works rests on treating the input as a parameter to be optimized, following [14], and on the local smoothness of the loss.
What would settle it
Train a classifier on a dataset with a strong spurious cue that correlates with class, such as MNIST digits pasted on colored backgrounds where each digit class has a fixed background color, and run the fill-in method on masked test images. If the generated prototypes reproduce the spurious background color or texture instead of the digit shape, the gradient update is reflecting superficial cues rather than true class structure, which would falsify the claim that the method reveals the model's prototypical structure.
Extended reading notes
Core claim
The paper's discovery is that a single gradient step of the classification loss with respect to input pixels—computed on a masked image—acts as the model's own suggestion for completing the missing regions. By thresholding this update and adding it back to the masked image, the authors obtain a visual answer to 'how would this model fill in the blanks?' Averaging over many random masks yields a prototypical image per class, and a progressive masking scheme can reconstruct an entire image from the model's updates. The authors test the fidelity of these prototypes by feeding them back to the model: the model classifies its own generated prototypical images with high accuracy (99% for random masking on MNIST, 93% for progressive). The conclusion is that the update direction encodes class-conditional structure the model has learned, making gradient-based fill-in a viable interpretability tool without training any extra network.
Load-bearing premise
The load-bearing premise is that the gradient of the classification loss with respect to input pixels, evaluated on a masked image, is a faithful proxy for the model's internal class-conditional prototype rather than being dominated by noise or low-level texture cues; if the gradient is noisy, the visualizations do not reveal learned structure.
Editorial extensions
If this is right
- The method gives a training-free way to visualize class prototypes for any differentiable vision classifier, without a separate generator or decoder.
- Averaged random-mask prototypes offer a global class-level explanation that can complement per-sample saliency maps.
- The progressive masking scheme tests whether a model can reconstruct a recognizable instance from partial context, probing its internal generative assumptions.
- Because the update is a gradient, the method works wherever the loss is differentiable with respect to input, including non-CNN architectures.
- The same fill-in question can be posed to classifiers in other modalities, such as text sentiment models, as the authors suggest.
Reading between the lines
- A practical use would be shortcut detection: if a model relies on background cues, the fill-in prototypes should reproduce those cues, letting practitioners see what the model thinks defines a class beyond the intended object.
- The averaged prototype over random masks may be more stable than a single activation-maximization image, suggesting it could serve as a cheap sanity check for whether a model has learned human-aligned class structure.
- The masking distribution itself is a probe design choice; different mask shapes or a curriculum of masks could reveal different aspects of the learned representation, such as local texture versus global shape.
- One testable extension: apply the same gradient-fill procedure to a vision transformer and compare its prototypes to those of a convolutional network to see whether the method exposes architecture-specific inductive biases.
Signed reviews
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper proposes an interpretability method for vision classifiers: given a masked input image, the method computes the gradient of the classification loss with respect to the input, thresholds and binarizes it, and superimposes the result on the masked image to visualize how the model 'fills in the blanks.' Two masking protocols are presented (random masks and progressive non-overlapping masks), and qualitative results are shown on MNIST and FashionMNIST with different classifier architectures. The authors also report a quantitative 'consistency' test in which the model classifies its own generated prototypical images.
Significance. If the approach worked as claimed, it would offer a lightweight, model-agnostic tool for visualizing class-level structure learned by vision models, without training a separate generator. The qualitative figures are suggestive, and the idea of using the model's own gradient to infer missing content is intuitive. However, the paper provides no rigorous quantitative evidence that the generated visualizations reflect true learned prototypes: the only numerical evaluation is circular, and the mathematical formulation of the core update is inconsistent with the described procedure. The paper's contribution as it stands is therefore not established.
major comments (5)
- [Section 3.2, Eq. (2)] The gradient in Eq. (2) is taken with respect to x of L(M(x), y), where M(x) is the element-wise product of a binary mask and the image. By the chain rule, this gradient vanishes at every masked pixel, because those entries do not affect the network output. Consequently, the application of the complement mask f_M in Eq. (4) selects entries that are identically zero, so the proposed update cannot fill any masked region. As written, the central derivation is inconsistent with the reported results. The authors must clarify whether they instead intend to differentiate with respect to the masked image z = M(x) and then transfer the gradient values to the masked positions, or to treat the masked pixels as free variables initialized to zero.
- [Section 4.3, Table 1] The quantitative evaluation is circular and not probative. The 'prototypical images' are generated by the same model f that is later asked to classify them. In the random-mask protocol of Section 4.1, 75% of each image consists of unmodified original pixels, and standard MNIST/FashionMNIST classifiers are highly robust to 25% random pixel dropout. The reported >93% accuracy is therefore expected regardless of whether the filled patches contain any meaningful structure. No zero-fill, noise-fill, wrong-class-fill, or independent-classifier controls are provided, so the experiments cannot distinguish 'f recognizes its own locally linear gradient directions' from 'the fill resembles the class prototype.' This is the only quantitative support for the paper's central claim, and it is not probative.
- [Section 4.1, Eq. (5)] The empirical average in Eq. (5) is computed only over masks for which the model predicts the correct class (the N* subset). This conditioning introduces a strong bias: the 'prototype' is selected from mask instances where the unmasked context already allowed a correct prediction, so the resulting average may simply reflect that context rather than any class-conditional feature the model has learned. The authors do not justify this selection, nor do they report results without it or compare against a random selection of masks.
- [Section 4.4] The method introduces several free hyperparameters—masking ratio η, patch size, gradient step size γ, binarization threshold, and the correctness filter in Eq. (5)—yet the paper provides no quantitative sensitivity analysis or principled procedure for setting them. The visual ablation in Figure 4 is anecdotal, and the acknowledgement that 'the hyper-parameters would have to be tuned depending on the dataset' (Section 4.4) is an admission that the method's behavior is not stable across settings without manual intervention.
- [Section 5] The concluding claim that the approach 'is able to generate visual outputs which resemble the structures that are conditioned on the masked image' is supported only by qualitative inspection. There is no quantitative measure of resemblance, no comparison to existing interpretability methods (e.g., activation maximization, saliency maps, or generative inpainting), and no human evaluation. The paper's central claim therefore rests on anecdote rather than evidence.
minor comments (5)
- [Throughout] There are numerous typos and grammatical errors, including 'intergrated' (abstract), 'Explainabe AI' (Index Terms), 'relavance' (Section 2.1), 'eventhough' (Section 2.1), 'interprit' (Section 2.2), and inconsistent capitalization. A careful proofread is needed.
- [Section 3.2 and 4.1] The notation is confusing: the same symbol x appears to refer to the original image, the current image, and the variable being optimized. For instance, in Eq. (2) x is updated, but in Section 4.1 x is the original image and M(x) is the masked version. The authors should adopt distinct notation (e.g., z for the masked/current image) to avoid ambiguity.
- [Section 4.1] The paper states that N is made sufficiently large so that N* > 10,000, but it does not report the actual values of N, the number of test images, or the distribution of masks. Reproducibility would be improved by specifying all experiment hyperparameters (γ, number of gradient steps, threshold values, model architectures, and training details).
- [Section 2.2] The related work on feature visualization is incomplete: recent methods such as concept-based explanations, textual explanations, and modern activation maximization with priors are not discussed. While not load-bearing, a more complete literature review would position the contribution better.
- [Section 4.3] The claimed 'consistency' is only measured for a single model per dataset. The paper would be stronger if it reported across multiple architectures, random seeds, and dataset splits, with error bars.
Circularity Check
Quantitative validation is self-referential: f is tested on images generated by f's own gradients, so Table 1 does not demonstrate learned class structure.
-
fitted input called prediction
[Section 4.3 / Table 1; construction in Section 3.2, Eq. (2)]
"We create a new test dataset with prototypical images generated from our method of visualisation. For each image, the label is the predicted label ˆy for the masked input. Afterwards, we test the model f to see if it can classify the prototypical images correctly... This confirms that our proposed prototypical image resembles the original class, both in the perspective of a human observer, and from the the perspective of the trained model f itself."
The 'prototypical images' in this test set are exactly M(x)+T(∆x), with ∆x = −γ∇_x L_θ(M(x), y) from Eq. (2). The filled pixels are therefore assembled from the direction in which f's own classification loss for the chosen label decreases. Re-feeding these images to f measures whether f recognizes inputs that were constructed to lower f's own loss toward that label; correct classification is largely forced by the construction. In the random-mask protocol, 75% of each image is untouched original content that MNIST/Fashion-MNIST classifiers already classify reliably, and no independent classifier, no zero-fill/noise-fill baseline, and no human study is reported. Table 1 cannot confirm resemblance to a learned class prototype because the test inputs are fitted to the very model being tested.
-
other
[Section 4.1, Eq. (5)]
"If f has learnt some prototypical structure from a particular class the expected value of the update should resemble the underlying structure of the class... From the N instances of M (x), we select the masks which result in a correct prediction of the class."
The 'expected update' is estimated only over masks on which f already predicts the class correctly. This conditions the averaged gradient on the model's own success, so the resulting 'prototypical image' is an artifact of filtering by f's correct predictions. It does not independently confirm that the update direction reflects a learned prototype; it ensures that only directions from already-successful predictions are averaged. This selection bias is built into the construction of the visual claim, not a check of it.
full rationale
The paper proposes a self-contained visualization recipe: mask an input, take one gradient step of the classifier's loss with respect to the input, binarize the masked-region update, and superimpose it on the masked image. The visualizations in Figures 2-6 are not circular by themselves; they are the output of a stated transformation of f and M. Circularity enters at the validation step. Section 4.3 constructs a 'test set' of prototypical images using this transformation and then asks the same model f to classify them. Because the filled pixels are generated by descending f's own classification loss toward the label (Eq. 2), f's high accuracy is largely a consequence of how the images were made, not independent evidence that the fill captures a class prototype. The random-mask protocol leaves 75% of each image unmodified, and no control condition (e.g., zero-fill, noise-fill, or wrong-class-gradient fill) or independent classifier is provided, so Table 1's 99% accuracy does not reveal whether the filled regions carry class structure. Additionally, Eq. (5) averages only over masks that already yield correct predictions, further conditioning the 'prototypical image' on f's own successes. No self-citation chain is load-bearing, and the method itself is not fitted to an external target; however, the sole quantitative test reduces to a self-consistency check, giving partial circularity (score 6).
Assumptions & free parameters
free parameters (6)
- masking ratio η =
0.25 (default)
- patch size =
2x2 pixels (default)
- gradient step size γ =
unspecified
- number of masks N =
unspecified; N* > 10,000
- binarization threshold in T =
unspecified
- selection filter for correct predictions =
N/A
assumptions (5)
- standard math The loss function L is locally smooth in input space, so a small gradient step lowers the loss and moves the image toward the class.
- domain assumption Input gradients of a trained classifier encode semantic class structure rather than low-level artifacts.
- ad hoc to paper The thresholding and binarization transform T preserves meaningful signal and removes noise.
- ad hoc to paper Averaging over masks that give correct predictions yields an unbiased estimate of the class prototype.
- domain assumption The method is model-agnostic and will generalize to architectures beyond the tested CNNs.
Cite this review
Pith. "Pith review of Fill in the blanks: Rethinking Interpretability in vision." pith.science (2026). https://pith.science/paper/WASIRW4D
@misc{pith2026241110273,
author = {Pith},
title = {Pith review of: Fill in the blanks: Rethinking Interpretability in vision},
year = {2026},
howpublished = {\url{https://pith.science/paper/WASIRW4D}},
note = {Machine review of arXiv:2411.10273}
}
read the original abstract
Model interpretability is a key challenge that has yet to align with the advancements observed in contemporary state-of-the-art deep learning models. In particular, deep learning aided vision tasks require interpretability, in order for their adoption in more specialized domains such as medical imaging. Although the field of explainable AI (XAI) developed methods for interpreting vision models along with early convolutional neural networks, recent XAI research has mainly focused on assigning attributes via saliency maps. As such, these methods are restricted to providing explanations at a sample level, and many explainability methods suffer from low adaptability across a wide range of vision models. In our work, we re-think vision-model explainability from a novel perspective, to probe the general input structure that a model has learnt during its training. To this end, we ask the question: "How would a vision model fill-in a masked-image". Experiments on standard vision datasets and pre-trained models reveal consistent patterns, and could be intergrated as an additional model-agnostic explainability tool in modern machine-learning platforms. The code will be available at \url{https://github.com/BoTZ-TND/FillingTheBlanks.git}
Reference graph
Works this paper leans on
-
[1]
INTRODUCTION The rapid advancements of Machine Learning (ML) and Deep Learning(DL) models show no signs of slowing down, with multimodal models further alluring researchers into pushing performance limits. However, the challenge of adopting DL architectures into critical applications such as medical imag- ing is largely due to our limited understanding of...
-
[2]
Introduce a mask-filling approach for visually interpret- ing vision models, as an alternative to existing genera- arXiv:2411.10273v1 [cs.CV] 15 Nov 2024 tive methods
work page Pith review arXiv 2024
-
[3]
Present different approaches of masking, and interpret the visual results on standard data sets
-
[4]
Demonstrate consistency of the visual patterns and present the effects of changing masking parameters
-
[5]
RELA TED WORK Our work aims to interpret learned features of an image clas- sifier based on a visual output. The conclusions on the model will be interpreted by the human, based on perceived patterns of the visual result. Therefore, we present a brief overview of literature that can be considered within XAI methods of ‘perceptive interpretability’. 2.1. S...
-
[6]
Marco Tulio Ribeiro, Sameer Singh, and Carlos Guestrin, ““why should i trust you?”: Explaining the predictions of any classifier,” Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining , 2016
work page 2016
-
[7]
METHODOLOGY 3.1. Rethinking interpretability Let us focus on the task of image classification, and re-visit how a human recalls from memory. When an image, say of the number ‘three’, is shown to a human, we associate that image with previous memories of the digit, and identify dis- tinctive features of the digit. However, this mental image of a digit is n...
-
[8]
How f gener- ates the prototypical class based on the given structure
The positive (red) and negative (blue) components of the update term fM (x) is shown superimposed with M (x). The thresholded update T (∆x) ‘fills in the blanks’ of the masked input. tilt, font...etc. Consequently, as observed by [21], searching for an image that maximized a single activation generates an image that a human would not perceive as a ‘three’...
Show all 33 references
-
[9]
Can the model recognise our visualisation of the ‘prototypical image’, as an instance of the original class ?
EXPERIMENTS (a) (b) Fig. 2: Visualising approximations of E[∆x] for digits of (a) MNIST[20] and (b) Fashion MNIST[22] data (section 4.1). Observe that the prototypical image of each class is visually convincing of distinctive features of that class. In this section, we present...
-
[10]
How would a vision model fill in masked-out patches?
is claimed to be applicable to any CNN-based vision model. We also note that the metrics that are visualized as saliency maps are derived from gradients that flow from a par- ticular node/activation of the network. Thereby, they are eas- ily implemented from existing framework...
-
[11]
As such, we propose to visualise how a pre- trained image classifier would ‘Fill in the blanks’ of a masked image
CONCLUSIONS AND FUTURE WORK In this work, we re-think the task of interpreting vision-based classifier models. As such, we propose to visualise how a pre- trained image classifier would ‘Fill in the blanks’ of a masked image. We show two approaches of masking, to demonstrate w...
-
[12]
Methods for interpreting and un- derstanding deep neural networks,
Gr ´egoire Montavon, Wojciech Samek, and Klaus- Robert M ¨uller, “Methods for interpreting and un- derstanding deep neural networks,” CoRR, vol. abs/1706.07979, 2017
2017 arXiv
-
[13]
Visualizing and understanding convolutional networks,
Matthew D. Zeiler and Rob Fergus, “Visualizing and understanding convolutional networks,” ArXiv, vol. abs/1311.2901, 2013
2013 arXiv
-
[14]
Under- standing deep image representations by inverting them,
Aravindh Mahendran and Andrea Vedaldi, “Under- standing deep image representations by inverting them,” in 2015 IEEE Conference on Computer Vision and Pat- tern Recognition (CVPR), 2014, pp. 5188–5196
2015
-
[15]
Understanding neural net- works through deep visualization,
Jason Yosinski, Jeff Clune, Anh Nguyen, Thomas Fuchs, and Hod Lipson, “Understanding neural net- works through deep visualization,” CoRR, 2015
2015
-
[16]
On pixel-wise explanations for non-linear classifier decisions by layer-wise relevance propagation.,
Bach S, Binder A, Montavon G, Klauschen F, M ¨uller K-R, and Samek W, “On pixel-wise explanations for non-linear classifier decisions by layer-wise relevance propagation.,” PLoS ONE, vol. 10, pp. 7, 2015
2015
-
[17]
Learning important features through propagating activation differences,
Avanti Shrikumar, Peyton Greenside, and Anshul Kun- daje, “Learning important features through propagating activation differences,” in International Conference on Machine Learning (ICML), 2017
2017
-
[18]
Ax- iomatic attribution for deep networks,
Mukund Sundararajan, Ankur Taly, and Qiqi Yan, “Ax- iomatic attribution for deep networks,” in International Conference on Machine Learning (ICML) , 2017
2017
-
[19]
Learning deep features for discriminative localization,
Bolei Zhou, Aditya Khosla, `Agata Lapedriza, Aude Oliva, and Antonio Torralba, “Learning deep features for discriminative localization,” 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , pp. 2921–2929, 2015
2016
-
[20]
Grad-cam: Visual explanations from deep net- works via gradient-based localization,
Ramprasaath R. Selvaraju, Abhishek Das, Ramakrishna Vedantam, Michael Cogswell, Devi Parikh, and Dhruv Batra, “Grad-cam: Visual explanations from deep net- works via gradient-based localization,” International Journal of Computer Vision , vol. 128, pp. 336 – 359, 2016
2016
-
[21]
Score-cam: Score-weighted visual explanations for convolutional neural networks,
Haofan Wang, Zifan Wang, Mengnan Du, Fan Yang, Zi- jian Zhang, Sirui Ding, Piotr (Peter) Mardziel, and Xia Hu, “Score-cam: Score-weighted visual explanations for convolutional neural networks,” 2020 IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion Workshops (CV...
2020
-
[22]
Ablation-cam: Visual explanations for deep con- volutional network via gradient-free localization,
Saurabh Satish Desai and H. G. Ramaswamy, “Ablation-cam: Visual explanations for deep con- volutional network via gradient-free localization,” 2020 IEEE Winter Conference on Applications of Computer Vision (WACV), pp. 972–980, 2020
2020
-
[23]
Axiom-based grad-cam: To- wards accurate visualization and explanation of cnns,
Ruigang Fu, Qingyong Hu, Xiaohu Dong, Yulan Guo, Yinghui Gao, and Biao Li, “Axiom-based grad-cam: To- wards accurate visualization and explanation of cnns,” ArXiv, vol. abs/2008.02312, 2020
2008 arXiv
-
[24]
Visualizing higher-layer features of a deep network,
D. Erhan, Yoshua Bengio, Aaron C. Courville, and Pas- cal Vincent, “Visualizing higher-layer features of a deep network,” 2009
2009
-
[25]
Deep inside convolutional networks: Visualising image classification models and saliency maps,
Karen Simonyan, Andrea Vedaldi, and Andrew Zisser- man, “Deep inside convolutional networks: Visualising image classification models and saliency maps,” CoRR, vol. abs/1312.6034, 2013
2013 arXiv
-
[26]
Striving for simplic- ity: The all convolutional net,
Jost Tobias Springenberg, Alexey Dosovitskiy, Thomas Brox, and Martin A. Riedmiller, “Striving for simplic- ity: The all convolutional net,” in International Confer- ence on Learning Representations (ICLR) , 2014
2014
-
[27]
Visualizing deep convolutional neural networks using natural pre- images,
Aravindh Mahendran and Andrea Vedaldi, “Visualizing deep convolutional neural networks using natural pre- images,” in International Journal of Computer Vision , 2015, pp. 233–255
2015
-
[28]
Synthesizing the pre- ferred inputs for neurons in neural networks via deep generator networks,
Anh M Nguyen, Alexey Dosovitskiy, Jason Yosinski, Thomas Brox, and Jeff Clune, “Synthesizing the pre- ferred inputs for neurons in neural networks via deep generator networks,” in Neural Information Processing Systems, 2016
2016
-
[29]
Plug and amp; play generative net- works: Conditional iterative generation of images in la- tent space,
A. Nguyen, J. Clune, Y . Bengio, A. Dosovitskiy, and J. Yosinski, “Plug and amp; play generative net- works: Conditional iterative generation of images in la- tent space,” in 2017 IEEE Conference on Computer Vi- sion and Pattern Recognition (CVPR) , 2017, pp. 3510– 3520
2017
-
[30]
Mnist handwritten digit database,
Yann LeCun, Corinna Cortes, and CJ Burges, “Mnist handwritten digit database,” ATT Labs [Online]. Avail- able: http://yann.lecun.com/exdb/mnist, vol. 2, 2010
2010
-
[31]
Feature visualization,
Chris Olah, Alexander Mordvintsev, and Ludwig Schubert, “Feature visualization,” Distill, 2017, https://distill.pub/2017/feature-visualization
2017
-
[32]
Fashion- mnist: a novel image dataset for benchmarking machine learning algorithms,
Han Xiao, Kashif Rasul, and Roland V ollgraf, “Fashion- mnist: a novel image dataset for benchmarking machine learning algorithms,” ArXiv, vol. abs/1708.07747, 2017
2017 arXiv
-
[33]
Deep residual learning for image recognition,
Kaiming He, X. Zhang, Shaoqing Ren, and Jian Sun, “Deep residual learning for image recognition,” in 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015
2016
Reviewed August 12, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.