Research map / Label-efficient and robust learning / Test-time adaptation
TIDE: Training Locally Interpretable Domain Generalization Models Enables Test-time Correction
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 TIDE, a single-source domain generalization method that trains a ResNet-18 with concept saliency alignment and local concept contrastive losses using diffusion-and-LLM-generated concept maps, plus iterative saliency-masking test-time correction.
Problem
Single-source domain generalization fails under semantic shifts (background, viewpoint) because models learn global features rather than domain-invariant local concepts; no datasets with per-class concept localization maps exist.
Approach
Automatically annotate per-class concept saliency maps by combining GPT-3.5 concept lists, Stable Diffusion cross-attention maps, and DIFT feature correspondence to transfer them to real DG images; then train with a saliency alignment loss and a local concept contrastive loss, and correct predictions at test time by iteratively masking salient regions until concept features match stored signatures
Key results
- TIDE reaches 80.02% average accuracy on PACS, versus 71.87% for the prior best (PromptD), with test-time correction.
- On VLCS TIDE averages 77.08%, a 13.37% gain over the second-best method per the paper.
- Ablation on PACS shows the LLCC loss raises average accuracy from 49.93% to 73.67%, and test-time correction adds a further 6.35 points to 80.02%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use TIDE instead of PromptD | PACS single-source domain generalization · PACS | average accuracy: +8.33% | Helped |
| Use TIDE instead of DSU-MAD | VLCS single-source domain generalization · VLCS | average accuracy: +13.37% | Helped |
| Use TIDE instead of NJPP | OfficeHome single-source domain generalization · OfficeHome | average accuracy: +16.16% | Helped |
| Use TIDE instead of NJPP | DomainNet single-source domain generalization · DomainNet | average accuracy: +8.84% | Helped |
| Add concept classification loss (ablation) | TIDE class-loss-only model · PACS | average accuracy: 49.35 -> 49.71 | No clear effect |
| Add concept saliency alignment loss (ablation) | TIDE with class and concept losses · PACS | average accuracy: 49.71 -> 49.93 | No clear effect |
| Add local concept contrastive loss (ablation) | TIDE with class, concept, and saliency alignment losses · PACS | average accuracy: 49.93 -> 73.67 | Helped |
| Add test-time correction (ablation) | TIDE without test-time correction · PACS | average accuracy: 73.67 -> 80.02 | Helped |
Surprising findings
- Adding the concept classification loss and saliency alignment loss alone barely changes accuracy (49.35 to 49.93 on PACS); the big gain comes from the contrastive loss and test-time correction.
Limitations
- Test-time correction is only invoked when concept signatures mismatch; in the PACS photo-to-sketch error analysis 27.8% of samples triggered correction and only 52.5% of those converged to the correct class.
- Concept discovery relies on a threshold tau and on GradCAM overlap from a ResNet-18 trained on the source domain.
- Evaluation uses only ResNet-18 and the SSDG setting; generalization to other backbones or fine-grained datasets is not shown.
Experimental setup
Datasets: PACS VLCS OfficeHome DomainNet
Benchmarks: PACS VLCS OfficeHome DomainNet
Metrics: classification accuracy
Baselines: ERM AugMix RSC MixStyle ABA PromptD NJPP Meta-Casual DSU-MAD P-RC
Builds on: ABA DIFT GradCAM DomainBed
Closest papers by meaning
- Multimodal Unsupervised Domain Generalization by Retrieving Across the Modality Gap (ICLR 2025 · similarity 0.87)
- Explaining Domain Shifts in Language: Concept Erasing for Interpretable Image Classification (CVPR 2025 · similarity 0.87)
- START: A Generalized State Space Model with Saliency-Driven Token-Aware Transformation (NeurIPS 2024 · similarity 0.86)
- PracticalDG: Perturbation Distillation on Vision-Language Models for Hybrid Domain Generalization (CVPR 2024 · similarity 0.86)
- Is Large-scale Pretraining the Secret to Good Domain Generalization? (ICLR 2025 · similarity 0.86)
- Reasoning-Driven Multimodal LLM for Domain Generalization (ICLR 2026 · similarity 0.86)
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.