Pith. sign in

REVIEW 3 major objections 5 minor 54 references

Neural Image Compression and Explanation

T0 review · 3 major / 5 minor · reviewed 2026-08-14 · deepseek-v4-flash

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

Pith's one-line read This paper claims that a single learned binary mask over image pixels, trained jointly with a CNN classifier under $L_0$ sparsity and smoothness constraints, both explains the classifier's prediction and compresses the image to about…

desk verdict NICE cleanly unifies learned saliency masks and mixed-resolution compression, but the headline compression result rests on an unfair finetuning comparison. read the letter →

arxiv 1908.08988 v2 pith:VTXYIVZU submitted 2019-08-09 cs.CV cs.LGstat.ML

Xiang Li , Shihao Ji This is my paper · ORCID
classification cs.CVcs.LGstat.ML
keywords ExplainableAIsparsitylearningdatacompressiondeepneuralnetworkssaliencymixed-resolutionimageshardconcreteestimator
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

NICE is an end-to-end pipeline that learns, for each input image, a sparse binary mask over pixels by optimizing the classifier's loss on the masked image plus an $L_0$ sparsity penalty and a smoothness term. The mask is the explanation: pixels kept at 1 are the ones whose presence most supports the final prediction, and the same mask turns the image into a mixed-resolution version, high-resolution on salient pixels and subsampled elsewhere. Because the mask is produced by a small generator network, new images get masks in a single forward pass, with no pixel-level annotations and no backpropagation at test time. On Caltech256 the paper reports that compressed images reach about $0.6\times$ the original PNG file size while classification accuracy drops only slightly (78.30% to 74.95% at block size 8). The payoff is one mechanism that simultaneously gives a decision explanation, a compressed artifact for storage or transmission, and fast inference for real-time systems.

What carries the argument

The load-bearing object is the stochastic binary mask $z \in \{0,1\}^P$ attached to every pixel, trained through the hard concrete gradient estimator, a reparameterization that stretches a binary concrete distribution and hard-sigmoids it into $[0,1]$, making the discrete mask differentiable. The generator network $G(x;\theta_g)$ outputs $\log\alpha$ for each pixel; sampling $z \sim q(z|\log\alpha)$ and minimizing $\mathcal{L} = \mathcal{L}_D + \lambda_1 \mathcal{L}_C + \lambda_2 \mathcal{L}_S$ lets gradients flow to the mask parameters, where $\mathcal{L}_D$ is the classifier's loss on the masked image, $\mathcal{L}_C$ is the expected number of nonzero mask entries, and $\mathcal{L}_S$ penalizes total variation between neighboring mask values. This machinery carries the argument because it turns the combinatorial problem of choosing a sparse pixel subset into a differentiable end-to-end optimization, and the same sampled mask feeds the mixed-resolution transform $\tilde{x} = x \odot \hat{z} + x_b \odot (1-\hat{z})$ that produces the compressed image.

What would settle it

Take a held-out set of images and their trained NICE masks, then run the explained classifier on the images with the highlighted pixels removed, replaced by the low-resolution background used at training. If the original predicted class survives on a large fraction of these counterfactual images, the mask is not marking the pixels that actually drive the prediction; it is marking only one sufficient subset, and the claim that the mask measures each pixel's influence on the final prediction fails.

Watch

Extended reading notes

Core claim

The central claim is that a single $L_0$-regularized stochastic binary mask, learned jointly with a CNN, captures the saliency of each pixel measured by its influence on the final prediction, and that this same mask yields a semantically compressed image that keeps classification accuracy nearly intact. Saliency is defined operationally: a pixel is important if removing it (masking it to a low-resolution background) hurts the classifier's loss, and the training objective combines that data loss with an expected-count penalty on nonzero mask entries and a smoothness penalty on neighboring mask values. The paper demonstrates the claim on MNIST, CIFAR10 and Caltech256, showing that masks concentrate on object regions, that randomizing the top-K masked pixels degrades accuracy far more than randomizing the bottom-K or random pixels, and that mixed-resolution images built from the masks compress to about $0.6\times$ the original file size (54 KB vs 87 KB at block size 8 on Caltech256) while retaining a similar classification accuracy, especially when the discriminator's top layers are finetuned.

Load-bearing premise

The load-bearing premise is that pixels whose preservation keeps the classifier accurate are the pixels the classifier actually uses for its decision; the training loss never verifies that the chosen pixels match the original CNN's reasoning, so a mask can be faithful to accuracy without being faithful to the model's decision process.

Editorial extensions

If this is right

  • Prediction, explanation, and compressed image come out of one forward pass, so a system can record why a decision was made and store or transmit the evidence at the same time.
  • The sparsity level is a tunable knob ($\lambda_1$), so applications can trade explanation conciseness against accuracy without retraining the architecture.
  • Finetuning the top layers of the classifier along with the mask generator improves the compression-accuracy tradeoff, a flexibility that backpropagation-based explanation baselines do not have.
  • Because inference needs no backpropagation, the method is fast enough for real-time deployment: about $23\times$ faster than Saliency Map, $16.5\times$ faster than CAM, and $2.8\times$ faster than RTIS on the reported GPU runs.
  • The mask generator transfers to new datasets for classes it has seen, so a trained generator can be applied to other high-resolution image collections without retraining.

Reading between the lines

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

  • Editorial inference: the objective allows a mask to be faithful to accuracy without being faithful to the original classifier's reasoning, because any sufficient subset of pixels that preserves the class scores well, including spurious cues; a stricter faithfulness test would check that keeping only the highlighted pixels reproduces the prediction and that removing them flips it.
  • Editorial inference: the same mask-plus-mixed-resolution trick applies to other input modalities, such as text tokens or genomic features, where a sparse 'keep these positions' mask doubles as a rationalization and as an input compression.
  • Editorial inference: pairing NICE masks with a learned lossy codec instead of PNG could push file sizes well below $0.6\times$, since the mask already removes semantic redundancy and the codec only needs to handle the remaining low-resolution background.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper proposes NICE, an end-to-end pipeline that trains a stochastic binary mask generator under L0 sparsity and smoothness constraints to explain CNN predictions and to produce mixed-resolution images for semantic compression. The mask is optimized with the hard-concrete gradient estimator, and at test time one forward pass through the generator yields the explanation and the mask needed for compression. Experiments on MNIST, CIFAR10, and Caltech256 compare NICE with Saliency Map, RTIS, and CAM for explanation quality, and with those methods plus downsampling for semantic compression, reporting a 1.6x compression rate at block size 8 with a 3.35% accuracy drop and about 23x faster inference than Saliency Map.

Significance. If the main claims are properly supported, NICE is a useful unified framework: it couples explanation and semantic compression in one differentiable pipeline, uses a principled L0 penalty with a scalable hard-concrete estimator, and ships code. The inductive mask generator and the speed advantage over backpropagation-based explainers are concrete strengths. However, the central compression superiority claim rests on an asymmetric evaluation protocol, and the explanation evaluation is close to a restatement of the training objective, so the significance of the reported numbers depends on additional controlled experiments.

major comments (3)
  1. [IV.C, Fig. 9] The headline semantic-compression comparison is not controlled. The text states that NICE-finetuned updates Conv-4, Conv-5, and the FC layers of ResNet18 on mixed-resolution images, while Saliency Map, RTIS, CAM, and downsampling are evaluated with the original fixed pretrained discriminator. The paper even notes that the baselines do not have the flexibility of finetuning their discriminators. Therefore the reported 1.6x compression at block size 8 with only a 3.35% accuracy drop can be attributed to classifier adaptation rather than to the quality of the sparse mask. To support the central claim, the authors should compare NICE-fixed against fixed-classifier baselines and also give the baselines the same finetuning treatment on their own mixed-resolution outputs; without this matched protocol the semantic-compression superiority claim is unverified.
  2. [III.A, Eq. (3); IV.B, Fig. 6] The explanation evaluation is closely aligned with the training objective, making it partly circular. The mask is trained by minimizing the classification loss of the masked image plus sparsity and smoothness terms, and Fig. 6 evaluates explanation quality by randomizing the top-K% mask pixels and measuring the accuracy drop. This measures whether the mask selects a sufficient subset of pixels, which is nearly what Eq. (3) optimizes; it does not independently establish that the mask corresponds to the pixels the original classifier actually uses, since any sufficient subset, including spurious cues, would score well. Additional validation, such as a human study, comparison with ground-truth segmentation masks, or perturbation tests that control for classifier re-adaptation, is needed before claiming that NICE explains the original CNN's decision process.
  3. [IV.C, text near Fig. 9] The compression rate figures are reported inconsistently. The abstract says the produced images achieve 'about 0.6x of original image file size,' while Section IV.C reports '1.6x compression rate (87KB vs. 54KB).' These are equivalent (54/87 is approximately 0.62), but the two phrasings should be reconciled to avoid confusion about whether the reported rate is a size ratio or a compression factor.
minor comments (5)
  1. [Abstract] The phrase 'a numerous of applications' is ungrammatical and should read 'numerous applications.'
  2. [Fig. 7 caption] There is a typo in 'Caltach256'; it should be 'Caltech256.'
  3. [Eq. (6)] In Eq. (6), the notation λLC(log α) is introduced before the regularization weight λ1 is defined in Eq. (11); the line should make clear that λ in the second term is λ1.
  4. [IV.C, Fig. 9] Fig. 9 reports file sizes and accuracies without error bars or repeated-trial statistics; adding standard deviations or confidence intervals would help assess the reliability of the reported 1.6x compression point.
  5. [IV.B, Fig. 6] The y-axis label 'Accuracy' should specify that this is the classification accuracy of the original pretrained classifier on the corrupted images, since the exact protocol matters for interpreting the curves.

Circularity Check

2 steps flagged · score 6.0 of 10

Explanation validation reuses the training objective, and the headline compression accuracy is obtained with a finetuned discriminator that the baselines were not allowed to adapt; central quantitative claims are partially circular.

  1. self definitional [Section III.A, Eq. (3); Section IV.B, Fig. 6]
    "We regard zi as our explanation to the prediction of h(xi;θ) and learn zi by minimizing the following L0-norm regularized loss function: R(θ,z)=... L(h(xi⊙zi;θ),yi)+λ||zi||0 ... To evaluate NICE's performance of identifying important pixels from an image, Fig. 6 demonstrates the evolution of classification accuracies on the Caltech256 test dataset when different percentages of pixels are filled with random values ... NICE identifies important pixels from images as randomizing their Top-K% values incurs a dramatic accuracy loss."

    The explanation is defined as the mask z that minimizes classification loss of the masked image (Eq. 3); a pixel is 'salient' exactly insofar as keeping it preserves the classifier's prediction. The post-hoc evaluation then validates the mask by corrupting the top-K% pixels and measuring classification accuracy loss. Corrupting the kept pixels is closely the complement of the training operation, so the evaluation metric is essentially the same quantity that Eq. (3) optimizes. The result is therefore a consistency check of the optimization, not an independent test that the selected pixels are the ones the original CNN uses; any sufficient subset trained under Eq. (3) would exhibit the same property.

  2. fitted input called prediction [Section IV.C, Fig. 9]
    "Discriminator-finetuned: similar to discriminator-fixed except that the top few layers of the discriminator θd are finetuned. In this case, the discriminator can adjust its parameters to improve its predictions on the mixed-resolution images, and thus higher accuracy and compression rate are expected. Note that due to their specific training methodologies, Saliency Map, RTIS and CAM do not have the flexibility of finetuning their discriminators ... When the block-size is 8, NICE-finetuned achieves a 1.6x compression rate (87KB vs. 54KB) with a small (3.35%) accuracy drop (78.30% vs. 74.95%)."

    The headline compression accuracy is produced under the Discriminator-finetuned protocol: the top layers of ResNet18 are trained on NICE's mixed-resolution images, so the reported 74.95% accuracy is the data term of the joint objective (11) after optimizing both the mask and θd, not an independent measure of what the original classifier can recognize. Baseline methods (Saliency Map, RTIS, CAM) are evaluated with the original fixed pretrained classifier, and the paper explicitly states they lack 'the flexibility of finetuning their discriminators.' The comparison therefore attributes to the mask an accuracy gain that is confounded with classifier adaptation. The 1.6x compression claim is not a controlled prediction of mask quality; it is a fitted value of the jointly trained discriminator.

full rationale

The core optimization machinery is not circular: Eqs. (3)-(11) define a self-contained training objective for a mask generator and a discriminator, and the hard-concrete estimator is cited from external work rather than from the authors' own prior claims. There are no load-bearing self-citations or imported uniqueness theorems. The circularity is located in the evaluation. First, the explanation claim is validated by the same accuracy-under-masking criterion used to train the mask, so the 'top-K% pixels hurt accuracy' result is a restatement of the training objective rather than an independent prediction. Second, the main compression result relies on a finetuned discriminator while the baselines use a fixed pretrained classifier, making the accuracy comparison asymmetric and the reported gain partially a property of the adapted classifier. The score is 6 rather than higher because NICE does train a genuine generator, the paper also reports the NICE-fixed protocol, and the circularity is in the evaluation design rather than in a derivation that is identical to its input by definition.

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

No new physical entities are introduced. The scientific assumptions are the five domain assumptions above; the manual hyperparameters lambda1, lambda2, b, and the adopted hard concrete constants are the main free parameters. The mask generator is a designed neural network component, not a postulated entity.

free parameters (4)
  • lambda1 (L0 penalty weight) = MNIST: 1, 10, 30; CIFAR10: 3; Caltech256: 5
    Tuned on validation set per dataset; it directly sets the sparsity of the explanation mask and therefore the compression/accuracy tradeoff. Different values are shown for MNIST, so the reported concise explanations depend on this choice.
  • lambda2 (smoothness penalty weight) = MNIST: 0; CIFAR10: 0.01; Caltech256: 0.01
    Chosen on validation to balance coherence and sparsity; affects the connectedness of masks and thus the visual quality of explanations.
  • Block size b = 16 for training; evaluated at 1, 2, 4, 8, 16, 32, 64
    Controls how aggressively the background is subsampled in the mixed-resolution image. It is the main knob for the compression vs. accuracy tradeoff and is not determined by the data.
  • Hard concrete constants beta, gamma, zeta = beta=2/3, gamma=-0.1, zeta=1.1 (from Louizos et al.)
    Adopted from prior work rather than tuned here, but they shape the relaxed mask values and the final thresholded masks, so the central results inherit them as hand-chosen constants.
assumptions (5)
  • domain assumption A mask that preserves the classifier's label under masking defines a valid explanation of that prediction.
    Eq. (3) and Eq. (11) use only classification loss plus sparsity/smoothness terms; no ground-truth relevance or independent attribution is used. If the classifier can be fooled into keeping the label with spurious pixels, the explanation can be unfaithful.
  • domain assumption The hard concrete relaxation is a valid surrogate for the intractable expectation over binary masks.
    Eqs. (6)-(8) replace Bernoulli sampling with a continuous hard-sigmoid relaxation. The paper adopts this from Louizos et al. without a formal guarantee that the relaxed optimum matches the original L0 objective.
  • domain assumption Pixel saliency can be modeled by independent Bernoulli gates per pixel plus a spatial smoothness penalty.
    Eq. (2) attaches one binary variable per pixel and Eq. (9) adds only local neighbor penalties; this cannot capture higher-order dependencies or hierarchical object structure.
  • domain assumption Replacing non-salient pixels with a block-averaged low-resolution background is an acceptable compression transform for the classifier.
    Eq. (13) defines the mixed-resolution image; the training and evaluation assume the classifier can still perform well on such images after fine-tuning, which is true on Caltech256 but not guaranteed for other distributions.
  • domain assumption The generator trained on one dataset assigns meaningful masks to new images from a related distribution.
    Section III.C turns the transductive per-image mask into an inductive generator G(x; theta_g). Transfer to ImageNet in Fig. 7 shows this only works for overlapping classes, so the generalization claim is class-dependent.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Neural Image Compression and Explanation." pith.science (2026). https://pith.science/paper/VTXYIVZU

@misc{pith2026190808988,
  author       = {Pith},
  title        = {Pith review of: Neural Image Compression and Explanation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VTXYIVZU}},
  note         = {Machine review of arXiv:1908.08988}
}
read the original abstract

Explaining the prediction of deep neural networks (DNNs) and semantic image compression are two active research areas of deep learning with a numerous of applications in decision-critical systems, such as surveillance cameras, drones and self-driving cars, where interpretable decision is critical and storage/network bandwidth is limited. In this paper, we propose a novel end-to-end Neural Image Compression and Explanation (NICE) framework that learns to (1) explain the predictions of convolutional neural networks (CNNs), and (2) subsequently compress the input images for efficient storage or transmission. Specifically, NICE generates a sparse mask over an input image by attaching a stochastic binary gate to each pixel of the image, whose parameters are learned through the interaction with the CNN classifier to be explained. The generated mask is able to capture the saliency of each pixel measured by its influence to the final prediction of CNN; it can also be used to produce a mixed-resolution image, where important pixels maintain their original high resolution and insignificant background pixels are subsampled to a low resolution. The produced images achieve a high compression rate (e.g., about 0.6x of original image file size), while retaining a similar classification accuracy. Extensive experiments across multiple image classification benchmarks demonstrate the superior performance of NICE compared to the state-of-the-art methods in terms of explanation quality and semantic image compression rate. Our code is available at: https://github.com/lxuniverse/NICE.

Figures

Figures reproduced from arXiv: 1908.08988 by the authors.

Figure 1
Figure 1. FIGURE 1 [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. FIGURE 2 [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. FIGURE 3 [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: FIGURE 4 [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: illustrates the comparison of NICE with Saliency Map [4], RTIS [9] and CAM [21] on the Caltech256 images. As we can see, our algorithm highlights the whole body of object as the explanation while Saliency Map typically identify edges or scattered pixels as the explanat…
Figure 7
Figure 7. Figure 7: illustrates some sample ImageNet images and their sparse masks. Images (a, b, c) are from the classes that included in Caltech256, while images (d, e) are from the classes that are not in Caltech256. It shows that for the classes that overlap with Caltech256, the gener…
Figure 6
Figure 6. Figure 6: FIGURE 6 [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 8
Figure 8. Figure 8: FIGURE 8 [PITH_FULL_IMAGE:figures/full_fig_p008_8.png]
Figure 9
Figure 9. Figure 9: shows the average file size of compressed images and the corresponding classification accuracy as a function of block size b. As we can see, when the block size in￾creases, the file size of the compressed images decreases (higher compression rate) and the classificatio…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

54 extracted references · 41 canonical work pages

  1. [1]

    Deep residual learning for image recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” in IEEE Conference on Computer Vision and Pattern Recog- nition (CVPR), pp. 770–778, 2016

    work page 2016

  2. [2]

    Bert: Pre-training of deep bidirectional transformers for language understanding,

    J. Devlin, M.-W. Chang, K. Lee, and K. Toutanova, “Bert: Pre-training of deep bidirectional transformers for language understanding,” arXiv preprint arXiv:1810.04805, 2018

    arXiv 2018

  3. [3]

    Exploring neu- ral transducers for end-to-end speech recognition,

    E. Battenberg, J. Chen, R. Child, A. Coates, Y . Gaur, Y . Li, H. Liu, S. Satheesh, D. Seetapun, A. Sriram, and Z. Zhu, “Exploring neu- ral transducers for end-to-end speech recognition,” arXiv preprint arXiv:1707.07413, 2017

    arXiv 2017

  4. [4]

    Deep inside convolutional networks: Visualising image classification models and saliency maps,

    K. Simonyan, A. Vedaldi, and A. Zisserman, “Deep inside convolutional networks: Visualising image classification models and saliency maps,” arXiv preprint arXiv:1312.6034, 2013

    arXiv 2013

  5. [5]

    Visualizing and understanding convolutional networks,

    M. D. Zeiler and R. Fergus, “Visualizing and understanding convolutional networks,” in European conference on computer vision (ECCV), 2014

    work page 2014

  6. [6]

    On pixel-wise explanations for non-linear classifier decisions by layer-wise relevance propagation,

    S. Bach, A. Binder, G. Montavon, F. Klauschen, K.-R. Müller, and W. Samek, “On pixel-wise explanations for non-linear classifier decisions by layer-wise relevance propagation,” PloS one, vol. 10, no. 7, 2015

    work page 2015

  7. [7]

    Devnet: A deep event network for multimedia event detection and evidence re- counting,

    C. Gan, N. Wang, Y . Yang, D.-Y . Yeung, and A. G. Hauptmann, “Devnet: A deep event network for multimedia event detection and evidence re- counting,” in IEEE Conference on Computer Vision and Pattern Recogni- tion (CVPR), pp. 2568–2577, 2015

    work page 2015

  8. [8]

    Why should i trust you?: Explaining the predictions of any classifier,

    M. T. Ribeiro, S. Singh, and C. Guestrin, “Why should i trust you?: Explaining the predictions of any classifier,” in Proceedings of the 22nd ACM SIGKDD international conference on knowledge discovery and data mining, pp. 1135–1144, 2016

    work page 2016

Show all 54 references
  1. [9]

    Real time image saliency for black box classifiers,

    P. Dabkowski and Y . Gal, “Real time image saliency for black box classifiers,” in NIPS, 2017

    2017

  2. [10]

    Interpretable explanations of black boxes by meaningful perturbation,

    R. C. Fong and A. Vedaldi, “Interpretable explanations of black boxes by meaningful perturbation,” in IEEE International Conference on Computer Vision (CVPR), pp. 3429–3437, 2017

    2017

  3. [11]

    Understanding neural networks through representation erasure,

    J. Li, W. Monroe, and D. Jurafsky, “Understanding neural networks through representation erasure,” arXiv preprint arXiv:1612.08220, 2016

    2016 arXiv

  4. [12]

    A unified approach to interpreting model predictions,

    S. M. Lundberg and S.-I. Lee, “A unified approach to interpreting model predictions,” in NIPS, 2017

    2017

  5. [13]

    Learning important features through propagating activation differences,

    A. Shrikumar, P. Greenside, and A. Kundaje, “Learning important features through propagating activation differences,” in ICML, 2017

    2017

  6. [14]

    Global model interpretation via recursive partitioning,

    C. Yang, A. Rangarajan, and S. Ranka, “Global model interpretation via recursive partitioning,” in IEEE International Conference on Data Science and Systems (DSS), 2018

    2018

  7. [15]

    Interpreting deep neural networks through variable importance,

    J. Ish-Horowicz, D. Udwin, S. Flaxman, S. Filippi, and L. Crawford, “Interpreting deep neural networks through variable importance,” arXiv preprint arXiv:1901.09839, 2019

    1901 arXiv

  8. [16]

    White paper: Cisco visual networking index: Forecast and trends, 2017–2022,

    “White paper: Cisco visual networking index: Forecast and trends, 2017–2022,” tech. rep., 2019

    2017

  9. [17]

    End-to-end optimized image compression,

    J. Ballé, V . Laparra, and E. P. Simoncelli, “End-to-end optimized image compression,” in ICLR, 2017

    2017

  10. [18]

    Semantic perceptual image compression using deep convolution networks,

    A. Prakash, N. Moran, S. Garber, A. Dilillo, and J. Storer, “Semantic perceptual image compression using deep convolution networks,” 2017 Data Compression Conference (DCC), Apr 2017

    2017

  11. [19]

    Improved lossy image compres- sion with priming and spatially adaptive bit rates for recurrent networks,

    N. Johnston, D. Vincent, D. Minnen, M. Covell, S. Singh, T. Chinen, S. Jin Hwang, J. Shor, and G. Toderici, “Improved lossy image compres- sion with priming and spatially adaptive bit rates for recurrent networks,” in CVPR, pp. 4385–4393, 2018

    2018

  12. [20]

    Neural multi-scale image compression,

    K. Nakanishi, S. ichi Maeda, T. Miyato, and D. Okanohara, “Neural multi-scale image compression,” in Asian Conference on Computer Vision (ACCV), 2018

    2018

  13. [21]

    Learning deep features for discriminative localization,

    B. Zhou, A. Khosla, A. Lapedriza, A. Oliva, and A. Torralba, “Learning deep features for discriminative localization,” in IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016

    2016

  14. [22]

    Grad-cam: Visual explanations from deep networks via gradient- based localization,

    R. R. Selvaraju, M. Cogswell, A. Das, R. Vedantam, D. Parikh, and D. Ba- tra, “Grad-cam: Visual explanations from deep networks via gradient- based localization,” in IEEE International Conference on Computer Vision (CVPR), pp. 618–626, 2017

    2017

  15. [23]

    Learning to explain: An information-theoretic perspective on model interpretation,

    J. Chen, L. Song, M. J. Wainwright, and M. I. Jordan, “Learning to explain: An information-theoretic perspective on model interpretation,” arXiv preprint arXiv:1802.07814, 2018

    2018 arXiv

  16. [24]

    Explaining a black-box us- ing deep variational information bottleneck approach,

    S. Bang, P. Xie, W. Wu, and E. Xing, “Explaining a black-box us- ing deep variational information bottleneck approach,” arXiv preprint arXiv:1902.06918, 2019

    1902 arXiv

  17. [25]

    Machine learning interpretability: A survey on methods and metrics,

    D. V . Carvalho, E. M. Pereira, and J. S. Cardoso, “Machine learning interpretability: A survey on methods and metrics,” Electronics, vol. 8, 2019

    2019

  18. [26]

    Fully convolutional networks for semantic segmentation,

    J. Long, E. Shelhamer, and T. Darrell, “Fully convolutional networks for semantic segmentation,” in Proceedings of the IEEE conference on computer vision and pattern recognition, 2015

    2015

  19. [27]

    Mask r-cnn,

    K. He, G. Gkioxari, P. Dollár, and R. Girshick, “Mask r-cnn,” in Proceed- ings of the IEEE international conference on computer vision, 2017

    2017

  20. [28]

    Encoder- decoder with atrous separable convolution for semantic image segmen- tation,

    L.-C. Chen, Y . Zhu, G. Papandreou, F. Schroff, and H. Adam, “Encoder- decoder with atrous separable convolution for semantic image segmen- tation,” in Proceedings of the European conference on computer vision (ECCV), 2018

    2018

  21. [29]

    Weakly-and semi-supervised learning of a deep convolutional network for semantic image segmentation,

    G. Papandreou, L.-C. Chen, K. P. Murphy, and A. L. Yuille, “Weakly-and semi-supervised learning of a deep convolutional network for semantic image segmentation,” in Proceedings of the IEEE international conference on computer vision, 2015

    2015

  22. [30]

    From image-level to pixel-level labeling with convolutional networks,

    P. O. Pinheiro and R. Collobert, “From image-level to pixel-level labeling with convolutional networks,” in Proceedings of the IEEE conference on computer vision and pattern recognition, 2015

    2015

  23. [31]

    What’s the point: Semantic segmentation with point supervision,

    A. Bearman, O. Russakovsky, V . Ferrari, and L. Fei-Fei, “What’s the point: Semantic segmentation with point supervision,” in European conference on computer vision, 2016

    2016

  24. [32]

    Unsupervised object segmentation by redrawing,

    M. Chen, T. Artières, and L. Denoyer, “Unsupervised object segmentation by redrawing,” in Advances in Neural Information Processing Systems (NIPS), 2019

    2019

  25. [33]

    Boxsup: Exploiting bounding boxes to supervise convolutional networks for semantic segmentation,

    J. Dai, K. He, and J. Sun, “Boxsup: Exploiting bounding boxes to supervise convolutional networks for semantic segmentation,” in Proceedings of the IEEE international conference on computer vision, 2015

    2015

  26. [34]

    Scribblesup: Scribble-supervised convolutional networks for semantic segmentation,

    D. Lin, J. Dai, J. Jia, K. He, and J. Sun, “Scribblesup: Scribble-supervised convolutional networks for semantic segmentation,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2016

    2016

  27. [35]

    The jpeg still picture compression standard,

    G. K. Wallace, “The jpeg still picture compression standard,” IEEE trans- actions on consumer electronics, vol. 38, no. 1, 1992

    1992

  28. [36]

    Sayood, Lossless compression handbook

    K. Sayood, Lossless compression handbook. Elsevier, 2002

    2002

  29. [37]

    End-to-end optimized image compression,

    J. Ballé, V . Laparra, and E. P. Simoncelli, “End-to-end optimized image compression,” arXiv preprint arXiv:1611.01704, 2016

    2016 arXiv

  30. [38]

    Variable rate image compression 10 VOLUME 4, 2016 X. Li et al.: Neural Image Compression and Explanation with recurrent neural networks,

    G. Toderici, S. M. O’Malley, S. J. Hwang, D. Vincent, D. Minnen, S. Baluja, M. Covell, and R. Sukthankar, “Variable rate image compression 10 VOLUME 4, 2016 X. Li et al.: Neural Image Compression and Explanation with recurrent neural networks,” in International Conference on L...

    2016

  31. [39]

    Full resolution image compression with recurrent neural networks,

    G. Toderici, D. Vincent, N. Johnston, S. Jin Hwang, D. Minnen, J. Shor, and M. Covell, “Full resolution image compression with recurrent neural networks,” in IEEE Conference on Computer Vision and Pattern Recogni- tion (CVPR), pp. 5306–5314, 2017

    2017

  32. [40]

    Lossy image compres- sion with compressive autoencoders,

    L. Theis, W. Shi, A. Cunningham, and F. Huszár, “Lossy image compres- sion with compressive autoencoders,” arXiv preprint arXiv:1703.00395, 2017

    2017 arXiv

  33. [41]

    Learning convolutional networks for content-weighted image compression,

    M. Li, W. Zuo, S. Gu, D. Zhao, and D. Zhang, “Learning convolutional networks for content-weighted image compression,” in IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018

    2018

  34. [42]

    Stochastic variational optimization,

    T. Bird, J. Kunze, and D. Barber, “Stochastic variational optimization,” arXiv preprint arXiv:1809.04855, 2018

    2018 arXiv

  35. [43]

    Simple statistical gradient-following algorithms for con- nectionist reinforcement learning,

    R. J. Williams, “Simple statistical gradient-following algorithms for con- nectionist reinforcement learning,” Machine Learning, vol. 8, pp. 229–256, May 1992

    1992

  36. [44]

    Categorical reparameterization with gumbel-softmax,

    E. Jang, S. Gu, and B. Poole, “Categorical reparameterization with gumbel-softmax,” in International Conference on Learning Representa- tions (ICLR), 2017

    2017

  37. [45]

    The concrete distribution: A continuous relaxation of discrete random variables,

    C. J. Maddison, A. Mnih, and Y . W. Teh, “The concrete distribution: A continuous relaxation of discrete random variables,” in International Conference on Learning Representations (ICLR), 2017

    2017

  38. [46]

    Rebar: Low-variance, unbiased gradient estimates for discrete latent variable models,

    G. Tucker, A. Mnih, C. J. Maddison, J. Lawson, and J. Sohl-Dickstein, “Rebar: Low-variance, unbiased gradient estimates for discrete latent variable models,” in NIPS, 2017

    2017

  39. [47]

    Backprop- agation through the void: Optimizing control variates for black-box gradi- ent estimation,

    W. Grathwohl, D. Choi, Y . Wu, G. Roeder, and D. Duvenaud, “Backprop- agation through the void: Optimizing control variates for black-box gradi- ent estimation,” in International Conference on Learning Representations (ICLR), 2018

    2018

  40. [48]

    Learning sparse neural net- works through l0 regularization,

    C. Louizos, M. Welling, and D. P. Kingma, “Learning sparse neural net- works through l0 regularization,” in International Conference on Learning Representations (ICLR), 2018

    2018

  41. [49]

    Gradient-based learning applied to document recognition,

    Y . LeCun, L. Bottou, Y . Bengio, P. Haffner, et al., “Gradient-based learning applied to document recognition,” Proceedings of the IEEE, vol. 86, no. 11, pp. 2278–2324, 1998

    1998

  42. [50]

    Learning multiple layers of features from tiny images,

    A. Krizhevsky, “Learning multiple layers of features from tiny images,” tech. rep., 2009

    2009

  43. [51]

    Caltech-256 object category dataset,

    G. Griffin, A. Holub, and P. Perona, “Caltech-256 object category dataset,” 2007

    2007

  44. [52]

    Adam: A method for stochastic optimization,

    D. P. Kingma and J. Ba, “Adam: A method for stochastic optimization,” in International Conference on Learning Representations (ICLR), 2015

    2015

  45. [53]

    Imagenet: A large-scale hierarchical image database,

    J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei, “Imagenet: A large-scale hierarchical image database,” in IEEE conference on computer vision and pattern recognition (CVPR), pp. 248–255, 2009

    2009

  46. [54]

    Real-time adaptive image compression,

    O. Rippel and L. Bourdev, “Real-time adaptive image compression,” in International Conference on Machine Learning (ICML), 2017. XIANG LI received the undergraduate degree in electrical engineering from Donghua University, China, in 2013, and the master degree in pattern recogn...

    2017

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.