mlsanityData to 7 Oct 2026

Why are Saliency Maps Noisy? Cause of and Solution to Noisy Saliency Maps

arXiv 1902.04893 ↗First posted 13 Feb 2019cs.LG cs.CV stat.MLcited by 34

Not on the map. Why?

Abstract (arXiv)

Saliency Map, the gradient of the score function with respect to the input, is the most basic technique for interpreting deep neural network decisions. However, saliency maps are often visually noisy. Although several hypotheses were proposed to account for this phenomenon, there are few works that provide rigorous analyses of noisy saliency maps. In this paper, we firstly propose a new hypothesis that noise may occur in saliency maps when irrelevant features pass through ReLU activation functions. Then, we propose Rectified Gradient, a method that alleviates this problem through layer-wise thresholding during backpropagation. Experiments with neural networks trained on CIFAR-10 and ImageNet showed effectiveness of our method and its superiority to other attribution methods.

Key facts. Cited by 34 papers in the map (data through 7 Oct 2026). First posted on arXiv 13 Feb 2019.

Builds on

No reference list yet.

Built on by

Cited by year (2026 to 7 Oct)

Citing papers per year: 2021: 6; 2022: 6; 2023: 4; 2024: 6; 2025: 7; 2026: 5.

34 papers.

Sort:

  1. 2021Do Input Gradients Highlight Discriminative Features?cited by 52
  2. 2024Gradient based Feature Attribution in Explainable AI: A Technical Reviewcited by 34
  3. 2021Surrogate Model-Based Explainability Methods for Point Cloud NNscited by 18
  4. 2022New Definitions and Evaluations for Saliency Methods: Staying Intrinsic, Complete and Soundcited by 10
  5. 2023Causal Analysis for Robust Interpretability of Neural Networkscited by 8
  6. 2022Fidelity of Interpretability Methods and Perturbation Artifacts in Neural Networkscited by 7
  7. 2024Structured Gradient-based Interpretations via Norm-Regularized Adversarial Trainingcited by 7
  8. 2021Evaluation of Interpretability for Deep Learning algorithms in EEG Emotion Recognition: A case study in Autismcited by 6
  9. 2022Deeply Explain CNN via Hierarchical Decompositioncited by 6
  10. 2025Uncertainty Propagation in XAI: A Comparison of Analytical and Empirical Estimatorscited by 6
  11. 2024Expected Grad-CAM: Towards gradient faithfulnesscited by 4
  12. 2022Guiding Visual Attention in Deep Convolutional Neural Networks Based on Human Eye Movementscited by 3
22 more
  1. 2023Saliency strikes back: How filtering out high frequencies improves white-box explanationscited by 3
  2. 2022Evaluation of importance estimators in deep learning classifiers for Computed Tomographycited by 2
  3. 2023On Pitfalls of textitRemOve-And-Retrain: Data Processing Inequality Perspectivecited by 2
  4. 2023Towards Robust Aspect-based Sentiment Analysis through Non-counterfactual Augmentationscited by 2
  5. 2024CoProNN: Concept-based Prototypical Nearest Neighbors for Explaining Vision Modelscited by 2
  6. 2025Attribution Explanations for Deep Neural Networks: A Theoretical Perspectivecited by 2
  7. 2025On the Complexity-Faithfulness Trade-off of Gradient-Based Explanationscited by 2
  8. 2025On Spectral Properties of Gradient-based Explanation Methodscited by 2
  9. 2022Understanding CNNs from excitationscited by 1
  10. 2024Bidirectional Emergent Language in Situated Environmentscited by 1
  11. 2024Gaussian Smoothing in Saliency Maps: The Stability-Fidelity Trade-Off in Neural Network Interpretabilitycited by 1
  12. 2025AutoSIGHT: Automatic Eye Tracking-based System for Immediate Grading of Human experTisecited by 1
  13. 2021Advances in Electron Microscopy with Deep Learningcited by 0
  14. 2021Attribution Mask: Filtering Out Irrelevant Features By Recursively Focusing Attention on Inputs of DNNscited by 0
  15. 2021Deep Learning Image Recognition for Non-imagescited by 0
  16. 2025Unifying Perplexing Behaviors in Modified BP Attributions through Alignment Perspectivecited by 0
  17. 2025Query Circuits: Explaining How Language Models Answer User Promptscited by 0
  18. 2026Explore the Ideology of Deep Learning in ENSO Forecastscited by 0
  19. 2026Parallelizing Node-Level Explainability in Graph Neural Networkscited by 0
  20. 2026Training for Trustworthy Saliency Maps: Adversarial Training Meets Feature-Map Smoothingcited by 0
  21. 2026Foundation models for discovering robust biomarkers of neurological disorders from dynamic functional connectivitycited by 0
  22. 2026Verifying the Linear Representation Hypothesis: How Interpretable Are Vision SAEs?cited by 0