Research map / Image generation / Fast diffusion sampling
Iterative Search Attribution for Deep Neural Networks
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes Iterative Search Attribution (ISA), a gradient ascent/descent search-based attribution method for DNN image classifiers that iteratively clips unimportant features with a scale parameter, and reports higher Insertion scores and lower INFD than IG, FIG, GIG, BIG, SM and AGI on ImageNet.
Problem
Gradient-based attribution methods for DNNs depend on reference points, adversarial sample quality, or gradient rules, and existing integration paths are under-explored, leading to inconsistent interpretability.
Approach
Treats attribution as an autoregressive iterative search that alternates gradient ascent and descent with sign-based, XOR-masked feature perturbations, normalizes and offsets attribution values per iteration (with a scale factor S) so later-removed features rank higher, and combines ascent and descent attributions into a single score.
Key results
- On Inception-v3, ISA reaches Insertion AUC 0.7346 versus 0.6256 for the next best (AGI), and on VGG16 ISA reaches INFD 0.56 versus 3.36 for AGI.
- ISA attains the highest Insertion score on all three models (Inception-v3, ResNet-50, VGG16) on 1000 randomly sampled ImageNet images.
- Using ISA with ResNet-50 gives Insertion AUC 0.6065 vs 0.5136 for AGI, but its Deletion score (0.0512) is slightly lower than GIG's best-in-class deletion on some models, showing deletion is not uniformly best.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use ISA instead of AGI | Inception-v3 · ImageNet 1000 samples | Insertion score (AUC): 0.6256 -> 0.7346 | Helped |
| Use ISA instead of AGI | ResNet-50 · ImageNet 1000 samples | Insertion score (AUC): 0.5136 -> 0.6065 | Helped |
| Use ISA instead of AGI | VGG16 · ImageNet 1000 samples | Insertion score (AUC): 0.4304 -> 0.5082 | Helped |
| Use ISA instead of GIG | Inception-v3 · ImageNet 1000 samples | Deletion score (AUC): 0.0343 -> 0.0523 | Hurt |
| Use ISA instead of GIG | ResNet-50 · ImageNet 1000 samples | Deletion score (AUC): 0.0191 -> 0.0512 | Hurt |
| Use ISA instead of GIG | VGG16 · ImageNet 1000 samples | Deletion score (AUC): 0.0173 -> 0.0423 | Hurt |
| Use ISA instead of AGI | Inception-v3 · ImageNet 1000 samples | INFD score: 4.98 -> 4.82 | Helped |
| Use ISA instead of AGI | ResNet-50 · ImageNet 1000 samples | INFD score: 0.84 -> 0.83 | Helped |
| Use ISA instead of AGI | VGG16 · ImageNet 1000 samples | INFD score: 3.36 -> 0.56 | Helped |
| Add gradient ascent and gradient descent rather than using either alone (ablation) | ISA on Inception-v3 · ImageNet | Insertion score: 0.7042 / 0.7055 -> 0.7346 | Helped |
| Add gradient ascent and gradient descent rather than using either alone (ablation) | ISA on Inception-v3 · ImageNet | Deletion score: 0.0715 / 0.0739 -> 0.0523 | Helped |
| Vary the number of features removed per iteration; step size 5000 gives the best reported result (ablation) | ISA on Inception-v3 · ImageNet | Insertion score | Mixed |
Surprising findings
- Deletion scores do not consistently favor ISA; GIG has the lowest deletion score on ResNet-50 and VGG16, and the authors argue deletion is unreliable for adversarial networks.
- Gradient ascent-only and descent-only variants perform much worse on Insertion than combined search, indicating the combination matters.
Limitations
- Evaluation limited to image classification on ImageNet with 1000 samples and three CNN-style models.
- Computational cost is slightly higher than AGI (480 vs 400 propagations), and features are removed in equal amounts per iteration.
- Hyperparameters (step size, learning rate, scale S) are tuned on the evaluation setup, and the Deletion metric is acknowledged as unreliable.
Experimental setup
Datasets: ImageNet
Benchmarks: ImageNet
Metrics: Insertion score Deletion score INFD score
Baselines: IG FastIG GuidedIG BIG SaliencyMap AGI
Builds on: Integrated Gradients Adversarial Gradient Integration Boundary-based Integrated Gradients Guided Integrated Gradients Saliency Map
Closest papers by meaning
- Towards better understanding of gradient-based attribution methods for Deep Neural Networks (ICLR 2018 · similarity 0.90)
- How to Probe: Simple Yet Effective Techniques for Improving Post-hoc Explanations (ICLR 2025 · similarity 0.86)
- MFABA: A More Faithful and Accelerated Boundary-Based Attribution Method for Deep Neural Networks (AAAI 2024 · similarity 0.85)
- Fast Data Attribution for Text-to-Image Models (NeurIPS 2025 · similarity 0.85)
- Integrated Decision Gradients: Compute Your Attributions Where the Model Makes Its Decision (AAAI 2024 · similarity 0.85)
- Benchmarking the Attribution Quality of Vision Models (NeurIPS 2024 · similarity 0.85)
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.