Research map / Optimization / Optimizers
Full-Gradient Representation for Neural Network Visualization
Suraj Srinivas, François Fleuret
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| remove absolute-value post-processing from FullGrad (ablation) | FullGrad · MNIST digit flipping (8 → 3) | Δ log-odds: 8.26 ± 21.44 -> 12.93 ± 18.20 | Improved |
| remove absolute-value post-processing from FullGrad (ablation) | FullGrad · MNIST pixel perturbation, RF = 0.5 | fractional output change: 0.06 ± 0.13 -> 0.19 ± 0.19 | Hurt |
| remove absolute-value post-processing from FullGrad (ablation) | FullGrad · MNIST pixel perturbation, RF = 0.7 | fractional output change: 0.08 ± 0.11 -> 0.34 ± 0.23 | Hurt |
| remove absolute-value post-processing from FullGrad (ablation) | FullGrad · MNIST pixel perturbation, RF = 0.9 | fractional output change: 0.55 ± 0.30 -> 0.63 ± 0.27 | Hurt |
| use FullGrad rather than IntegratedGrad | saliency maps · MNIST pixel perturbation, RF = 0.5 | fractional output change: 0 ± 0 vs 0.06 ± 0.13 | Hurt |
| use FullGrad rather than IntegratedGrad | saliency maps · MNIST pixel perturbation, RF = 0.7 | fractional output change: 0.004 ± 0.06 vs 0.08 ± 0.11 | Hurt |
| use FullGrad rather than IntegratedGrad | saliency maps · MNIST pixel perturbation, RF = 0.9 | fractional output change: 0.44 ± 0.33 vs 0.55 ± 0.30 | Hurt |
| aggregate input-gradients with neuron bias-gradients | FullGrad · ImageNet 2012 validation pixel perturbation | absolute fractional output change | Improved |
| aggregate input-gradients with neuron bias-gradients | FullGrad · CIFAR100 remove-and-retrain | accuracy after retraining | Improved |
Closest papers by meaning
- A Novel Visual Interpretability for Deep Neural Networks by Optimizing Activation Maps with Perturbation (AAAI 2021 · similarity 0.87)
- Visualization of Supervised and Self-Supervised Neural Networks via Attribution Guided Factorization (AAAI 2021 · similarity 0.86)
- Structured Gradient-based Interpretations via Norm-Regularized Adversarial Training (CVPR 2024 · similarity 0.86)
- Rethinking Interpretation: Input-Agnostic Saliency Mapping of Deep Visual Classifiers (AAAI 2023 · similarity 0.85)
- Unlearning-based Neural Interpretations (ICLR 2025 · similarity 0.83)
- Explaining the Implicit Neural Canvas: Connecting Pixels to Neurons by Tracing their Contributions (CVPR 2024 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
