Research map / Label-efficient and robust learning / Recognition and long-tail
Unlearning-based Neural Interpretations
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 UNI, which computes a learned, unlearning-based baseline for path-attribution (e.g., Integrated Gradients) by unlearning salient features in model space and matching activations in input space, yielding more faithful, robust attributions than static baselines.
Problem
Gradient-based path attribution methods such as Integrated Gradients depend on a baseline input representing absence of features; static baselines (black, blurred, noised) inject colour, texture and frequency biases and high-curvature paths make attributions noisy and fragile.
Approach
Replace static baselines with an adaptive baseline found by first-order unlearning of the model (ascent in loss along a steepest-ascent direction in weight space), then optimising a perturbation of the input so its class confidence matches the unlearned model's confidence on the original input; the resulting baseline defines a monotonic, low-curvature path for path attribution.
Key results
- UNI achieves the best MuFidelity in 5 of 6 architectures (e.g., ResNet-18 0.12 vs 0.06 for IG).
- UNI attains the highest insertion AUC across all six models (e.g., ViT-B_16 0.71, ResNet-18 0.64 vs 0.26 for IG on ResNet-18).
- Under FGSM attack, UNI keeps Spearman correlation around 0.27-0.32 across models, higher than IG, BlurIG, SG and DeepLIFT.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use UNI instead of the strongest listed baseline | ImageNet-1K attribution on ResNet-18 · ImageNet-1K | MuFidelity: .10 -> .12 | Helped |
| Use UNI instead of the strongest listed baseline | ImageNet-1K attribution on EfficientNet-v2-small · ImageNet-1K | MuFidelity: .06 ± .045 vs .06 ± .046 | No clear effect |
| Use UNI instead of GIG | ImageNet-1K attribution on ConvNeXt-Tiny · ImageNet-1K | MuFidelity: .18 -> .16 | Hurt |
| Use UNI instead of the strongest listed baseline | ImageNet-1K attribution on VGG-16-bn · ImageNet-1K | MuFidelity: .14 -> .18 | Helped |
| Use UNI instead of the strongest listed baseline | ImageNet-1K attribution on ViT-B-16 · ImageNet-1K | MuFidelity: .14 -> .15 | Helped |
| Use UNI instead of the strongest listed baseline | ImageNet-1K attribution on Swin-T-Tiny · ImageNet-1K | MuFidelity: .13 ± .102 vs .13 ± .100 | No clear effect |
| Use UNI instead of the strongest listed baseline | ImageNet-1K attribution across backbones · ImageNet-1K | Deletion AUC: .06 vs .08; .19 vs .18; .11 vs .17; .08 vs .10; .14 vs .13; .13 vs .22 | Mixed |
| Use UNI instead of the strongest listed baseline | ImageNet-1K attribution across backbones · ImageNet-1K | Insertion AUC: .56 -> .64; .51 -> .64; .56 -> .63; .47 -> .56; .59 -> .71; .63 -> .68 | Helped |
| Use UNI instead of IG | ImageNet attribution robustness across backbones · ImageNet-1K | Spearman correlation under FGSM: .088 -> .271; .009 -> .302; .010 -> .292; .143 -> .290; .018 -> .319; .088 -> .271 | Helped |
| Use UNI instead of Random baseline | Pythia-1b-v0 attribution patching · 100 counterfactual prompts | L2 distance to activation patching: 6.64 -> 3.12 | Helped |
| Use UNI instead of Random baseline | GPT2-medium attribution patching · 100 counterfactual prompts | L2 distance to activation patching: 35.00 -> 15.26 | Helped |
| Use UNI instead of Random baseline | Llama-3.2-1B attribution patching · 100 counterfactual prompts | L2 distance to activation patching: 10.17 -> 5.25 | Helped |
Surprising findings
- Static baselines such as black or blurred images inject colour, texture and frequency biases that are not present in the model itself, causing attribution errors even on corrupted inputs.
Limitations
- Gains on deletion AUC are not uniform; EfficientNet and ViT results show UNI not best on all metrics (e.g., AGI best on EfficientNet and ViT deletion AUC).
- NLP extension is preliminary, with no fine-tuning of hyperparameters and evaluation via L2 distance to activation patching on 100 prompts.
- Only first-order, sample-wise unlearning is explored; higher-level concept unlearning and second-order unlearning are left for future work.
Experimental setup
Datasets: ImageNet-1K ImageNet-C
Benchmarks: ImageNet-1K ImageNet-C
Metrics: MuFidelity deletion AUC insertion AUC Spearman correlation top-1000 pixel intersection LPIPS L2 distance to activation patching
Baselines: Integrated Gradients BlurIG GIG AGI GBP DeepLIFT Simple Gradients
Builds on: Integrated Gradients DeepLIFT BlurIG Expected Gradients activation patching
Closest papers by meaning
- Ascent Fails to Forget (NeurIPS 2025 · similarity 0.86)
- The BatchNorm Illusion: Diagnosing Normalization Artifacts in Machine Unlearning Evaluation (NeurIPS 2026 · similarity 0.85)
- Unified Gradient-Based Machine Unlearning with Remain Geometry Enhancement (NeurIPS 2024 · similarity 0.85)
- LibraGrad: Balancing Gradient Flow for Universally Better Vision Transformer Attributions (CVPR 2025 · similarity 0.85)
- DePass: Unified Feature Attributing by Simple Decomposed Forward Pass (NeurIPS 2025 · similarity 0.85)
- From Dormant to Deleted: Tamper-Resistant Unlearning Through Weight-Space Regularization (NeurIPS 2025 · 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.