Research map / Continual learning / Class-incremental
TALON: Test-time Adaptive Learning for On-the-Fly Category Discovery
Yanan Wu, Yuhan Yan, Tailai Chen, Zhixiang Chi, ZiZhang Wu, Yi Jin, Yang Wang, Zhenbo Li
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 |
|---|---|---|---|
| add margin-aware logit calibration (MLC) in offline training (ablation) | TALON baseline · Stanford Cars (All) | accuracy: 47.8 -> 49.0 | Improved |
| add MLC (ablation) | TALON baseline · CUB-200-2011 (All) | accuracy: 44.5 -> 45.7 | Improved |
| add semantic-aware prototype update (TTA-P) on top of MLC (ablation) | TALON · Stanford Cars (All) | accuracy: 49.0 -> 52.7 | Improved |
| add encoder adaptation (TTA-M) alone without MLC (ablation) | TALON baseline · CUB-200-2011 (All) | accuracy: 44.5 -> 44.5 | No clear effect |
| add encoder adaptation (TTA-M) on top of MLC (ablation) | TALON · Stanford Cars (All) | accuracy: 49.0 -> 52.1 | Improved |
| add both prototype and encoder adaptation on top of MLC (full TALON) (ablation) | TALON · Stanford Cars (All) | accuracy: 52.7 -> 53.5 | Improved |
| add both prototype and encoder adaptation on top of MLC (full TALON) (ablation) | TALON · CUB-200-2011 (All) | accuracy: 46.7 -> 45.5 | Hurt |
| replace TALON adaptation with TENT (ablation) | baseline+MLC · Stanford Cars (All) | accuracy: 49.0 -> 48.1 | Hurt |
| replace TALON adaptation with OSTTA (ablation) | baseline+MLC · Stanford Cars (All) | accuracy: 49.0 -> 47.2 | Hurt |
| use hash-free continuous prototypes instead of hash codes (vs PHE-64bit) | OCD · CUB-200-2011 (Strict-Hungarian) | estimated number of classes (true 200): 493 -> 153 | Improved |
| use hash-free continuous prototypes instead of hash codes (vs SMILE-64bit) | OCD · Stanford Cars (Strict-Hungarian) | estimated number of classes (true 196): 4788 -> 299 | Improved |
| use margin-aware logit calibration to shrink sample-to-prototype angle (ablation) | baseline features · Oxford Pets | mean angle to class prototype (degrees): 64.55 -> 35.83 | Improved |
Closest papers by meaning
- Learn to Categorize or Categorize to Learn? Self-Coding for Generalized Category Discovery (NeurIPS 2023 · similarity 0.90)
- On-the-Fly Category Discovery (CVPR 2023 · similarity 0.90)
- Prototypical Hash Encoding for On-the-Fly Fine-Grained Category Discovery (NeurIPS 2024 · similarity 0.89)
- Adaptive Discovering and Merging for Incremental Novel Class Discovery (AAAI 2024 · similarity 0.86)
- Prior-Constrained Association Learning for Fine-Grained Generalized Category Discovery (AAAI 2025 · similarity 0.85)
- Continual Generalized Category Discovery: Learning and Forgetting from a Bayesian Perspective (ICML 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.
