Research map / Label-efficient and robust learning / Domain adaptation and generalization
HiLo: A Learning Framework for Generalized Category Discovery Robust to Domain Shifts
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.
Introduces HiLo, a ViT-based generalized category discovery method robust to domain shifts, combining mutual-information-minimized low/high-level feature disentanglement, PatchMix contrastive learning, and domain-curriculum sampling, evaluated on DomainNet and a new corrupted SSB-C benchmark.
Problem
Generalized category discovery assumes unlabelled images share the labelled set's domain; in practice unlabelled data can contain covariate shifts and new classes, which existing GCD methods handle poorly.
Approach
Extract domain features from an early ViT layer and semantic features from the last layer, minimize their mutual information via a Jensen-Shannon estimator, apply PatchMix-style embedding mixup adapted to contrastive and clustering losses with attention-based semantic weights, and gradually increase sampling of unseen-domain unlabelled samples.
Key results
- On DomainNet Real+Painting, HiLo reaches 42.1 All ACC on Painting vs 34.5 for SimGCD and 64.4 vs 61.3 on Real.
- On SSB-C corrupted splits, HiLo reports 52.0 All ACC on CUB-C corrupted vs 28.8 for SimGCD.
- Ablation: using deep-only or shallow-only features for the two heads drops Real All ACC to 28.2 and 10.1, showing layer-wise disentanglement matters.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use HiLo instead of SimGCD | GCD with DINO ViT-B/16 · DomainNet Real+Painting, Painting | ACC All: 34.5 -> 42.1 | Helped |
| use HiLo instead of SimGCD | GCD with DINO ViT-B/16 · CUB-C Corrupted | ACC All: 28.8 -> 52.0 | Helped |
| use HiLo instead of SimGCD | GCD with DINO ViT-B/16 · Scars-C Corrupted | ACC All: 22.1 -> 35.6 | Helped |
| use HiLo instead of SimGCD | GCD with DINO ViT-B/16 · FGVC-C Corrupted | ACC All: 22.3 -> 31.2 | Helped |
| add original PatchMix (ablation) | SimGCD · DomainNet Real+Painting, Painting | ACC All: 34.5 -> 34.8 | Helped |
| adapt PatchMix for contrastive learning (ablation) | SimGCD · DomainNet Real+Painting, Painting | ACC All: 34.5 -> 36.6 | Helped |
| add disentangled domain and semantic features (ablation) | SimGCD · DomainNet Real+Painting, Painting | ACC All: 34.5 -> 35.6 | Helped |
| add curriculum sampling (ablation) | SimGCD · DomainNet Real+Painting, Painting | ACC All: 34.5 -> 38.4 | Helped |
| use deep features for both domain and semantic heads (ablation) | HiLo · DomainNet Real+Painting, Painting | ACC All: 42.1 -> 13.6 | Hurt |
| use shallow features for both domain and semantic heads (ablation) | HiLo · DomainNet Real+Painting, Painting | ACC All: 42.1 -> 5.7 | Hurt |
| add self-distillation for the domain head (ablation) | HiLo · DomainNet Real+Painting, Painting | ACC All: 42.1 -> 40.2 | Hurt |
| replace PatchMix with Mixup (ablation) | HiLo · DomainNet Real+Painting, Painting | ACC All: 42.1 -> 39.8 | Hurt |
Surprising findings
- Naively adding the UDA method PatchMix to SimGCD gives only marginal gains, and UDA techniques such as MCC and NWD yield only small improvements over SimGCD.
Limitations
- Relies on semi-supervised k-means over DINO domain features to estimate unseen domains, with no access to domain labels.
- SSB-C corruptions are synthetic rather than real-world shifts, and multi-domain DomainNet experiments were run on subsampled data.
Experimental setup
Datasets: DomainNet CUB Stanford Cars FGVC-Aircraft ImageNet-C corruptions
Benchmarks: DomainNet SSB-C (CUB-C, Scars-C, FGVC-C)
Metrics: clustering accuracy (ACC) All/Old/New ACC
Baselines: SimGCD GCD ORCA RankStats+ UNO+ UniOT MCC NWD EFDM SFA
Builds on: SimGCD PatchMix DINO ImageNet-C
Closest papers by meaning
- Generalized Category Discovery under Domain Shift: A Frequency Domain Perspective (NeurIPS 2025 · similarity 0.91)
- HIDISC: A Hyperbolic Framework for Domain Generalization with Generalized Category Discovery (NeurIPS 2025 · similarity 0.91)
- SEAL: Semantic-Aware Hierarchical Learning for Generalized Category Discovery (NeurIPS 2025 · similarity 0.90)
- When Domain Generalization meets Generalized Category Discovery: An Adaptive Task-Arithmetic Driven Approach (CVPR 2025 · similarity 0.90)
- Hyperbolic Category Discovery (CVPR 2025 · similarity 0.89)
- Transfer and Alignment Network for Generalized Category Discovery (AAAI 2024 · similarity 0.89)
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.