Pith. sign in

REVIEW 3 major objections 6 minor 31 references

Quantized and Interpretable Learning Scheme for Deep Neural Networks in Classification Task

T0 review · 3 major / 6 minor · reviewed 2026-08-11 · deepseek-v4-flash

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

Pith's one-line read Saliency-guided training and PACT quantization together preserve accuracy while making quantized models interpretable.

desk verdict Near-parity accuracy holds up; the efficiency and interpretability claims don't. read the letter →

arxiv 2412.03915 v1 pith:Z3G3NFGA submitted 2024-12-05 cs.LG cs.CV

Alireza Maleki , Mahsa Lavaei , Mohsen Bagheritabar , Salar Beigzad , Zahra Abadi This is my paper · ORCID
classification cs.LGcs.CV
keywords saliency-guidedtrainingquantization-awarePACTmodelinterpretabilityneuralnetworkquantizationCNNclassificationMNISTCIFAR-10
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper aims to establish that a neural network can be both quantized and interpretable without sacrificing classification accuracy. It combines saliency-guided training (SGT), which masks low-gradient input features during training, with PACT-based quantization-aware training, and tests this combination on MNIST and CIFAR-10 using a ResNet-20 architecture. The authors report that the combined approach achieves slightly higher accuracy than quantization alone (99.35% vs. 99.23% on MNIST; 76.12% vs. 75.55% on CIFAR-10) and produces saliency maps they argue are more focused. If true, this would make low-precision models for edge devices more trustworthy, since their decisions could be explained with the same quality as full-precision models.

What carries the argument

The central object is the combined loss $\mathcal{L} = L(f_\theta(X), y) + \lambda D_{\mathrm{KL}}(f_\theta(X) \| f_\theta(\tilde{X}))$, where $\tilde{X}$ is the input with the bottom $k$ features removed according to their input-gradient magnitude, used together with PACT's learnable clipping parameter $\alpha$ in a quantization-aware training loop. SGT provides the interpretability mechanism by forcing predictions to be invariant under removal of low-saliency features, while PACT provides the efficiency mechanism by quantizing activations and weights to 8 bits on MNIST and 4 bits on CIFAR-10. The dynamic optimization of $\alpha$ is what lets the model tune how aggressively activations are clipped during training.

What would settle it

Compare the accuracy-drop curve under random feature masks with the curve under gradient-ranked masks for the SGT+PACT model; if the two curves match, the masking test does not measure learned saliency and the interpretability claim is falsified.

Watch

Extended reading notes

Core claim

The central claim is that saliency-guided training and quantization-aware training are complementary: SGT teaches the network to rely on a compact set of salient features by minimizing the KL divergence between outputs on the original input and a version with low-gradient features masked, while PACT learns a per-layer clipping level $\alpha$ that keeps activations and weights at low precision. Combined, the method preserves classification accuracy (slightly improving over quantization alone) and yields models whose saliency maps are sharper. As evidence, the paper shows that when a larger fraction of low-gradient features is removed, the SGT-trained quantized model loses accuracy faster than a naively quantized model, which the authors interpret as the model having learned more salient features.

Load-bearing premise

The interpretability claim rests on the premise that a steeper accuracy drop when input features are masked is a valid measure of interpretability; the efficiency claim rests on the premise that lower bitwidths automatically mean lower resource use, since no latency or memory measurements are provided.

Editorial extensions

If this is right

  • If the combined approach works, low-precision models intended for phones and embedded devices can be trained with interpretability built in, so explanations do not have to be computed separately at deployment.
  • The reported accuracy parity suggests that saliency-guided regularization can offset some of the accuracy loss normally caused by aggressive quantization, implying the two techniques are compatible rather than competing.
  • The masking-ratio accuracy-drop test provides a cheap, training-time diagnostic for whether a quantized network relies on a focused set of features, which practitioners could adopt as a standard audit.
  • The dynamic $\alpha$ schedule indicates that the clipping level can be tuned together with the saliency objective, potentially removing the need to hand-tune quantization ranges.

Reading between the lines

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

  • The steeper accuracy drop under masking may be partly an artifact of the training objective itself, since SGT explicitly trains the model to be sensitive to the masked features; without a control that trains with random masks, the interpretability claim is confounded.
  • The efficiency claim is stated from bitwidth reduction alone; actual gains in latency, energy, or memory on a target device remain to be demonstrated, and future work could measure them directly.
  • The same recipe could be tested on tasks with ground-truth object locations, such as detection or segmentation, where saliency quality can be scored against human annotations instead of gradient-based proxies.
  • Quantization noise and saliency masking may interact positively as a form of regularization; this suggests a testable hypothesis that the SGT regularizer acts as a stabilizer for low-bitwidth training.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The paper proposes combining Saliency-Guided Training (SGT) with PACT-based quantization-aware training to obtain deep neural network classifiers that are, according to the abstract, both resource-efficient and interpretable without compromising accuracy. The method trains a ResNet-20 model by masking the k lowest-gradient input features, adding a KL-divergence term that penalizes divergence between predictions on original and masked inputs, and using PACT to quantize activations and weights during training. Experiments are reported on MNIST and CIFAR-10. Table II reports accuracies of 99.35% and 76.12% for the proposed method versus 99.23% and 75.55% for a PACT-only baseline. Figure 3 shows accuracy drops under increasing input masking as evidence of interpretability, and Figures 1-2 show the evolution of the PACT clipping parameter. The conclusion claims that the quantized models maintain near-equivalent accuracy while resource consumption is limited and interpretability is enhanced.

Significance. If fully substantiated, the paper would address a real practical need: quantized models for edge deployment whose explanations remain trustworthy. The combination of SGT with PACT is a reasonable and relatively simple empirical proposal, and the accuracy numbers in Table II are consistent with the modest claim that adding SGT does not hurt accuracy relative to a PACT-only baseline. However, the paper's central contribution is empirical, and at present only the accuracy-parity sub-claim has direct supporting measurements. The efficiency claim is supported only by an unquantified assertion in Section IX, and the interpretability claim rests on a single masking experiment whose outcome is closely tied to the training objective itself. The paper provides no code, no variance/seed information, and no direct saliency-map comparison despite the abstract promising one. These gaps prevent the headline conclusion from being accepted as demonstrated.

major comments (3)
  1. [Abstract and Section IX] The headline claim that the models are 'significantly more efficient' is not supported by any direct measurement. The paper reports no latency, memory footprint, FLOP/s, parameter count, energy consumption, or inference-speed comparison. Section IX only states that 'their resource consumption is limited,' which is not a quantitative result. To support the efficiency half of the central claim, the authors should report concrete efficiency metrics for the quantized models versus the full-precision and PACT baselines, with hardware and measurement details.
  2. [Section V, Algorithm 2, and Section VIII-C, Figure 3] The interpretability evidence is circular with respect to the training objective. The loss in Algorithm 2 and Section V is L = CE + lambda * D_KL(f_theta(X) || f_theta(M_k(X))), where M_k masks the k lowest-gradient features. The training procedure explicitly optimizes the model so that masking low-saliency features does not change the output, which is exactly the behavior measured in Figure 3: a steeper accuracy drop when salient features are masked. This makes Figure 3 a check that the training objective was optimized, not an independent demonstration that the learned saliency maps are more meaningful or more aligned with human-judged relevance. The interpretability claim needs additional evidence, such as quantitative saliency-map quality metrics, localization benchmarks, or human evaluation, and the paper should directly compare saliency maps as the abstract promises.
  3. [Table II and Section VIII-A] The accuracy comparison is under-specified. Table II reports only two columns, 'PACT' and 'Ours,' with no standard deviation, number of seeds, or test-set details, and the PACT baseline is not described (architecture, training hyperparameters, quantization bit widths, or whether it also uses SGT are not stated). There is also no full-precision baseline in the table, despite the text claiming that quantized models maintain near-equivalent accuracy to standard models. The authors should report mean and variance over multiple runs, describe the baseline precisely, and include the full-precision accuracy for both datasets.
minor comments (6)
  1. [Section I] The text contains an unremoved editorial instruction: 'Here's a refined version of your text with improved flow and clarity:' followed by the rewritten paragraph. This is clearly an artifact of the writing process and must be removed.
  2. [Algorithm 1 and Algorithm 2] Both algorithms contain a duplicated loop header 'for i = 1 to epochs do' with no matching indentation for the second occurrence. This makes the pseudocode ambiguous and should be corrected.
  3. [Section VII, Table I] The table caption reads 'Training Hyperparameters for Saliency-Guided Training with PACT Quantization and CIFAR-10 values,' which is grammatically incomplete and unclear. The caption should state clearly which hyperparameters apply to which dataset.
  4. [References] Reference [24] is cited both as 'mixed-precision quantization [24]' in Section I and as the PACT paper by Choi et al. later in the same section, but the reference list entry is only for PACT. Please disambiguate the citations or add the correct reference for mixed-precision quantization.
  5. [Section VIII-C] The caption of Figure 3 says 'Accuracy drop comparison across different models on MNIST dataset,' but the text in Section VIII-C discusses 'models with lower bitwidths (higher quantization)' experiencing sharper accuracy drops. It is unclear which models and bitwidths are actually plotted, and whether the figure shows CIFAR-10 results anywhere. Please clarify the experimental setup and the set of curves in the figure.
  6. [Section VII-A] The paper says the models are trained on 'ResNet-20' but gives no architecture details such as the number of parameters, the specific ResNet variant, or the input preprocessing. Adding these details would improve reproducibility.

Circularity Check

1 steps flagged · score 6.0 of 10

The interpretability claim rests on a masking-accuracy-drop metric that the SGT loss directly optimizes, making the main interpretability evidence circular; accuracy and efficiency claims are not themselves circular.

  1. self definitional [Algorithm 2 (Section VI) loss; Section VIII-C / Figure 3 caption; Section IX conclusion]
    "Algorithm 2 defines the training loss as: 'Li = L(yorig, y) + λDKL (yorig∥ymasked)'. Figure 3 caption states: 'our approach's accuracy drops more as the masking percentage increases. This is an illustration that the model has learned more salient features during the training.'"

    The DKL term in the SGT loss directly trains the network to keep its output distribution close when the k lowest-gradient input features are masked. The paper's only evidence for improved interpretability is Figure 3's accuracy-drop curve under feature masking, and the conclusion (Section IX) reads the steeper drop as 'the model is more aware and interpretable considering the input.' But that steeper drop is the behavioral signature the training loss was explicitly designed to produce: a model optimized to minimize DKL(fθ(X)||fθ(Mk(X))) is constructed to be sensitive to which features are masked. Thus the interpretability 'result' is equivalent to the training objective, not an independent confirmation that the saliency maps are more meaningful or aligned with human judgment.

full rationale

The accuracy comparison (Table II) is an independent, non-circular result: SGT+PACT numbers are compared against PACT baselines and the loss does not directly determine the final test accuracy value. The efficiency claim ('significantly more efficient') is asserted without latency, memory, FLOP, or energy measurements, which is a missing-evidence problem rather than a circularity. The self-citations (e.g., [25], [30]) appear in related-work context and are not used to justify the central derivation; SGT itself is attributed to Ismail et al. [19] and PACT to Choi et al. [24]. The one load-bearing circularity is the masking-sensitivity evidence for interpretability: Figure 3's accuracy-drop metric is the behavioral consequence of Algorithm 2's KL term, so the paper's main interpretability claim reduces to its own training objective. Score 6 reflects that one central component is circular while the accuracy-parity result remains independent.

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

The central claims rest entirely on published components (SGT and PACT) plus hand-set hyperparameters (masking ratio 50%, lambda, bit widths, learning rates). No new entities are introduced. The interpretability evidence assumes a masking-sensitivity proxy that is entangled with the SGT loss itself, and the efficiency claim assumes quantization gives efficiency without any measurement. Everything load-bearing in this paper is either inherited from prior work or chosen by hand.

free parameters (4)
  • PACT clipping parameter alpha = learned during training, final values not reported
    Learnable clipping threshold that defines activation quantization in the PACT scheme; the paper's quantization equations depend on it, so the reported accuracies are conditional on its fitted values and on the alpha-schedule used.
  • Saliency masking ratio k = 50% of features masked
    Chosen by hand (Table I); the SGT training dynamics and the Figure 3 masking curves depend on this choice, and no sensitivity analysis is given.
  • KL regularization weight lambda = 0.1 (MNIST), 0.05 (CIFAR-10)
    Hand-chosen hyperparameter balancing cross-entropy and the KL divergence that drives the interpretability behavior (Table I); the claimed interpretability gain is not tested across lambda values.
  • Quantization bit width = 8 bits (MNIST), 4 bits (CIFAR-10)
    A hand-set precision level (Table I); accuracy and the strength of the efficiency claim depend on this choice, and the two datasets use different precisions without justification.
assumptions (5)
  • domain assumption SGT improves interpretability: masking low-gradient features during training yields clearer and more reliable saliency maps.
    Inherited from Ismail et al. [19]; the paper's interpretability claim assumes this holds for quantized models without re-validating it on low-precision networks.
  • domain assumption Accuracy drop under input masking is a valid proxy for interpretability.
    Section VIII-C and Figure 3 interpret a steeper accuracy drop as evidence the model 'is more aware and interpretable'; no external validation (human study, faithfulness metrics, sanity checks) is provided.
  • domain assumption PACT maintains near-full-precision accuracy at low bit widths.
    Taken from Choi et al. [24]; the baseline comparison and the accuracy-parity claim presuppose this known result rather than re-deriving it.
  • domain assumption Quantization implies resource efficiency without direct measurement.
    The paper claims 'significantly more efficient' models but reports no FLOP, latency, memory, or energy numbers, treating the efficiency benefit of quantization as self-evident.
  • domain assumption ResNet-20 on MNIST and CIFAR-10 is representative enough to support the general claims in the abstract and conclusion.
    The conclusion generalizes to diverse architectures and datasets while only one architecture and two small benchmarks are tested, with no ablation across architectures.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Quantized and Interpretable Learning Scheme for Deep Neural Networks in Classification Task." pith.science (2026). https://pith.science/paper/Z3G3NFGA

@misc{pith2026241203915,
  author       = {Pith},
  title        = {Pith review of: Quantized and Interpretable Learning Scheme for Deep Neural Networks in Classification Task},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/Z3G3NFGA}},
  note         = {Machine review of arXiv:2412.03915}
}
read the original abstract

Deep learning techniques have proven highly effective in image classification, but their deployment in resourceconstrained environments remains challenging due to high computational demands. Furthermore, their interpretability is of high importance which demands even more available resources. In this work, we introduce an approach that combines saliency-guided training with quantization techniques to create an interpretable and resource-efficient model without compromising accuracy. We utilize Parameterized Clipping Activation (PACT) to perform quantization-aware training, specifically targeting activations and weights to optimize precision while minimizing resource usage. Concurrently, saliency-guided training is employed to enhance interpretability by iteratively masking features with low gradient values, leading to more focused and meaningful saliency maps. This training procedure helps in mitigating noisy gradients and yields models that provide clearer, more interpretable insights into their decision-making processes. To evaluate the impact of our approach, we conduct experiments using famous Convolutional Neural Networks (CNN) architecture on the MNIST and CIFAR-10 benchmark datasets as two popular datasets. We compare the saliency maps generated by standard and quantized models to assess the influence of quantization on both interpretability and classification accuracy. Our results demonstrate that the combined use of saliency-guided training and PACT-based quantization not only maintains classification performance but also produces models that are significantly more efficient and interpretable, making them suitable for deployment in resource-limited settings.

Figures

Figures reproduced from arXiv: 2412.03915 by the authors.

Figure 1
Figure 1. α Optimization for MNIST [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. α Optimization for CIFAR10 throughout training, the model finds the best value to balance the tradeoff between accuracy and quantization [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Accuracy drop comparison across different models on MNIST dataset [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

31 extracted references · 19 canonical work pages

  1. [1]

    Caruana, R., Lou, Y ., Gehrke, J., Koch, P., Sturm, M., & Elhadad, N. (2015). Intelligible models for healthcare: Predicting pneumonia risk and hospital 30-day readmission. In Proceedings of the 21th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (pp. 1721-1730)

    work page 2015

  2. [2]

    Li, X., Xie, L., Wang, J., & Yuille, A. L. (2018). Tell me where to look: Guided attention inference network. arXiv preprint arXiv:1802.10171

    work page Pith review arXiv 2018

  3. [3]

    & Padoy, N

    Hassanpour, J., Srivastav, V ., Mutter, D. & Padoy, N. Overcoming Di- mensional Collapse in Self-supervised Contrastive Learning for Medical Image Segmentation. ArXiv Preprint arXiv:2402.14611 . (2024)

    arXiv 2024

  4. [4]

    & Sasan, A

    Karkehabadi, A., Homayoun, H. & Sasan, A. FFCL: Forward-Forward Net with Cortical Loops, Training and Inference on Edge Without Backpropogation. Proceedings Of The Great Lakes Symposium On VLSI

  5. [5]

    R., Cogswell, M., Das, A., Vedantam, R., Parikh, D., & Batra, D

    Selvaraju, R. R., Cogswell, M., Das, A., Vedantam, R., Parikh, D., & Batra, D. (2017). Grad-CAM: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE International Conference on Computer Vision (pp. 618-626)

    work page 2017

  6. [6]

    Shrikumar, A., Greenside, P., & Kundaje, A. (2017). Learning important features through propagating activation differences. In Proceedings of the 34th International Conference on Machine Learning-Volume 70 (pp. 3145-3153)

    work page 2017

  7. [7]

    J., & Yu, B

    Singh, S., Murdoch, W. J., & Yu, B. (2017). Hide-and-seek saliency. arXiv preprint arXiv:1704.04289

    arXiv 2017

  8. [8]

    J., Sch ¨utt, K

    Kindermans, P. J., Sch ¨utt, K. T., Alber, M., M ¨uller, K. R., Erhan, D., Kim, B., & D ¨ahne, S. (2016). Investigating the influence of noise and distractors on the interpretation of neural networks. arXiv preprint arXiv:1611.07270

    arXiv 2016

Show all 31 references
  1. [9]

    Smilkov, D., Thorat, N., Kim, B., Vi ´egas, F., & Wattenberg, M. (2017). SmoothGrad: removing noise by adding noise. arXiv preprint arXiv:1706.03825

    2017 arXiv

  2. [10]

    Sundararajan, M., Taly, A., & Yan, Q. (2017). Axiomatic attribution for deep networks. In Proceedings of the 34th International Conference on Machine Learning-Volume 70 (pp. 3319-3328)

    2017

  3. [11]

    R., & Samek, W

    Bach, S., Binder, A., Montavon, G., Klauschen, F., M ¨uller, K. R., & Samek, W. (2015). On pixel-wise explanations for non-linear classi- fier decisions by layer-wise relevance propagation. PloS one , 10(7), e0130140

    2015

  4. [12]

    Adebayo, J., Gilmer, J., Muelly, M., Goodfellow, I., Hardt, M., & Kim, B. (2018). Sanity checks for saliency maps. In Advances in Neural Information Processing Systems (pp. 9505-9515)

    2018

  5. [13]

    Ghorbani, A., Abid, A., & Zou, J. (2019). Interpretation of neural networks is fragile. In Proceedings of the AAAI Conference on Artificial Intelligence (V ol. 33, pp. 3681-3688)

    2019

  6. [14]

    Hubara, I., Courbariaux, M., Soudry, D., El-Yaniv, R., & Bengio, Y . (2017). Quantized neural networks: Training neural networks with low precision weights and activations. The Journal of Machine Learning Research, 18(1), 6869-6898

    2017

  7. [15]

    Nagel, M., van Baalen, M., Blankevoort, T., & Welling, M. (2019). Data-free quantization through weight equalization and bias correction. In Proceedings of the IEEE/CVF International Conference on Computer Vision (pp. 1325-1334)

    2019

  8. [16]

    & Maknoon, R

    Pour, M., Zare, N. & Maknoon, R. Urban flood resilience assessment & stormwater management (case study: District 6 of Tehran). International Journal of Disaster Recovery and Business Continuity , 12(3), 215-234 (2024)

    2024

  9. [17]

    Han, S., Mao, H., & Dally, W. J. (2016). Deep compression: Compress- ing deep neural networks with pruning, trained quantization and huffman coding. arXiv preprint arXiv:1510.00149

    2016 arXiv

  10. [18]

    Y . Weng, S. Sun, L. He, and J. Lin, ”An overview of quanti- zation techniques for neural network compression,” arXiv preprint arXiv:2301.00388, 2023

    2023 arXiv

  11. [19]

    & Feizi, S

    Ismail, A., Corrada Bravo, H. & Feizi, S. Improving deep learning inter- pretability by saliency guided training. Advances In Neural Information Processing Systems. 34 pp. 26726-26739 (2021)

    2021

  12. [20]

    & Choi, K

    Chakraborty, S., Alam, M., Dey, S. & Choi, K. Interpretability of deep learning models: A survey of results. ArXiv Preprint arXiv:1707.09616. (2017)

    2017 arXiv

  13. [21]

    & Theiler, J

    Perkins, S., Lacker, L. & Theiler, J. Grafting: Fast, incremental feature selection by gradient descent in function space. Proceedings Of The 20th International Conference On Machine Learning (ICML-03). pp. 592-599 (2003)

    2003

  14. [22]

    & Hinton, G

    Frosst, N. & Hinton, G. Distilling a neural network into a soft decision tree. Proceedings Of The 20th International Conference On Machine Learning (ICML-17). pp. 1920-1930 (2017)

    2017

  15. [23]

    & Doshi-Velez, F

    Ross, A., Hughes, M. & Doshi-Velez, F. Right for the right rea- sons: Training differentiable models by constraining their explanations. Proceedings Of The 26th International Joint Conference On Artificial Intelligence (IJCAI). pp. 2662-2670 (2017)

    2017

  16. [24]

    Y . Choi, M. El-Khamy, and J. Lee, ”PACT: Parameterized clipping acti- vation for quantized neural networks,” arXiv preprint arXiv:1805.06085, 2018

    2018 arXiv

  17. [25]

    & Sasan, A

    Karkehabadi, A., Homayoun, H. & Sasan, A. SMOOT: Saliency guided mask optimized online training. 2024 IEEE 17th Dallas Circuits And Systems Conference (DCAS) . pp. 1-6 (2024)

    2024

  18. [26]

    & Burges, C

    LeCun, Y ., Cortes, C. & Burges, C. MNIST handwritten digit database. ATT Labs [Online]. Available: Http://yann. Lecun. Com/exdb/mnist . 2 (2010)

    2010

  19. [27]

    & Hinton, G

    Krizhevsky, A. & Hinton, G. Learning multiple layers of features from tiny images. (University of Toronto,2009)

    2009

  20. [28]

    Novel Saliency Analysis for the Forward Forward Algo- rithm

    Bakhshi, M. Novel Saliency Analysis for the Forward Forward Algo- rithm. ArXiv Preprint arXiv:2409.15365 . (2024)

    2024 arXiv

  21. [29]

    & Others Captum: A unified and generic model interpretability library for PyTorch

    Kokhlikyan, N., Miglani, V ., Martin, M., Wang, E., Reynolds, J., Melnikov, A., Lunghi, L., Reblitz-Richardson, O., Taylor, G., Cox, R. & Others Captum: A unified and generic model interpretability library for PyTorch. ArXiv Preprint arXiv:2009.07896 . (2020)

    2020 arXiv

  22. [30]

    & Sasan, A

    Karkehabadi, A., Latibari, B., Homayoun, H. & Sasan, A. HLGM: A Novel Methodology For Improving Model Accuracy Using Saliency- Guided High and Low Gradient Masking. The 14th International Con- ference On Information Science And Technology

  23. [31]

    & Zisserman, A

    Simonyan, K., Vedaldi, A. & Zisserman, A. Deep inside convolutional networks: Visualising image classification models and saliency maps. ArXiv Preprint arXiv:1312.6034 . (2013)

    2013 arXiv

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.