Research map / Vision-language and multimodal / CLIP and zero-shot
Advancing Analytic Class-Incremental Learning through Vision-Language Calibration
Binyu Zhao, Wei ZHANG, Xingrui Yu, Zhaonian Zou, Ivor Tsang
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 frozen CLIP branch (dual-branch) to the ViT-Adapter analytic classifier (ablation) | VILA · Cars (T=10) | last-task accuracy AT: 68.27 -> 84.13 | Improved |
| add unified geometric calibration (UGC) on top of dual-branch (ablation) | VILA · Cars (T=10) | last-task accuracy AT: 84.13 -> 90.45 | Improved |
| add candidate semantic enhancement (CSE) on top of DB+UGC (ablation) | VILA · Cars (T=10) | last-task accuracy AT: 90.45 -> 90.58 | Improved |
| add CSE on top of DB+UGC (ablation) | VILA · Aircraft (T=10) | last-task accuracy AT: 67.24 -> 67.26 | No clear effect |
| initialize analytic CIL with LoRA on the first task (ViT-LoRA) (ablation) | ViT analytic CIL · CIFAR (20 tasks) | last-task accuracy AT: 90.62 (10 tasks) -> 3.36 (20 tasks) | Hurt |
| use frozen CLIP visual weights directly in analytic CIL (ablation) | analytic CIL · ImageNet-R | last-task accuracy: 12.68% | Hurt |
| replace ViT-Adapter-only pilot with adapters over LoRA as specialized branch (ablation) | ViT analytic CIL · CIFAR (20 tasks) | last-task accuracy AT: 3.36 -> 89.39 | Improved |
| use adapter plus frozen CLIP (A+C) instead of adapter plus frozen ViT (A+V) (ablation) | VILA · Aircraft and ImageNet-R (T=20) | average accuracy | Improved |
| fuse all features (adapter, ViT and CLIP) (ablation) | VILA · Aircraft and ImageNet-R (T=20) | average accuracy | Hurt |
| replace frozen CLIP with frozen DINOv2 in DB+UGC (ablation) | VILA · Aircraft (T=10) | last-task accuracy AT: 67.24 -> 75.46 | Improved |
| replace frozen CLIP with frozen DINOv2 in DB+UGC (ablation) | VILA · SUN (T=30) | last-task accuracy AT: 83.60 -> 81.65 | Hurt |
| increase candidate size K in CSE (ablation) | VILA · CSE sensitivity | accuracy: saturates at K = 5, < 0.1% fluctuation | No clear effect |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- TUNA — TUNA: Taming Unified Visual Representations for Native Unified Multimodal Models (CVPR 2026) · better on 4 of 4 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- On the Power of Statistics in Class-Incremental Learning with Pretrained Models (ICML 2026 · similarity 0.90)
- Navigating Semantic Drift in Task-Agnostic Class-Incremental Learning (ICML 2025 · similarity 0.89)
- Advancing Cross-domain Discriminability in Continual Learning of Vision-Language Models (NeurIPS 2024 · similarity 0.88)
- Enhancing Few-Shot Class-Incremental Learning via Training-Free Bi-Level Modality Calibration (CVPR 2025 · similarity 0.88)
- AnaCP: Toward Upper-Bound Continual Learning via Analytic Contrastive Projection (NeurIPS 2025 · similarity 0.88)
- T-CIL: Temperature Scaling using Adversarial Perturbation for Calibration in Class-Incremental Learning (CVPR 2025 · similarity 0.87)
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.
