Research map / Bandits and online learning / Multi-armed bandits
Adaptive Data Augmentation with Multi-armed Bandit: Sample-Efficient Embedding Calibration for Implicit Pattern Recognition
Minxue Tang, Yangyang Yu, Aolin Ding, Maziyar Baran Pouyan, Taha Belkhouja, Yujia Bao
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 |
|---|---|---|---|
| replace random augmentation with MAB-based adaptive augmentation (ablation) | CLIP-VIT-Large calibration · CUB200 (zero-shot) | accuracy: 64.96% -> 68.60% | Improved |
| replace random augmentation with MAB-based adaptive augmentation (ablation) | OpenAI-text-embedding-3-small calibration · TREC (zero-shot) | accuracy: 57.56% -> 61.63% | Improved |
| replace random augmentation with MAB-based adaptive augmentation (ablation) | OpenAI-text-embedding-3-small calibration · Forbidden Question Set (few-shot) | accuracy: 84.62% -> 89.85% | Improved |
| replace random augmentation with MAB-based adaptive augmentation (ablation) | Voyage-multimodal-3 calibration · OxfordPets (zero-shot) | accuracy: 85.05% -> 86.34% | Improved |
| add random synthetic augmentation to calibration with only the initial set (ablation) | OpenAI-text-embedding-3-small calibration · MultiWD (zero-shot) | accuracy: 50.66% -> 56.83% | Improved |
| add adaptive augmentation to calibration with only the initial set (ablation) | QWen-3-emb-06b calibration · TREC (few-shot) | accuracy: 51.02% -> 62.50% | Improved |
| calibrate CLIP with light-weight calibrators and ADAMAB | CLIP-VIT-Large · CUB200 (zero-shot) | accuracy: 33.18% -> 68.60% | Improved |
| calibrate Voyage-multimodal-3 with ADAMAB | Voyage-multimodal-3 · Flowers102 (zero-shot) | accuracy: 49.13% -> 92.33% | Improved |
| calibration with only the initial set on few-shot descriptions | QWen-3-emb-06b · MultiWD (few-shot) | accuracy: 49.34% -> 44.93% | Hurt |
| add few-shot examples to in-context prompt | GPT-4o-mini · Forbidden Question Set | accuracy: 80.31% -> 76.00% | Hurt |
| compare ADAMAB-calibrated CLIP with Gemini2.0-Flash-Lite | CLIP-VIT-Large calibration · OxfordPets (zero-shot) | accuracy: 93.32% -> 93.20% | No clear effect |
| use exploration alpha > 0 instead of alpha = 0 (greedy) (ablation) | ADAMAB · all datasets | accuracy | Improved |
Closest papers by meaning
- Gram-Calibrated Anchoring for Class-Incremental Learning (NeurIPS 2026 · similarity 0.83)
- Adam with Bandit Sampling for Deep Learning (NeurIPS 2020 · similarity 0.83)
- Improving Few-Shot Generalization by Exploring and Exploiting Auxiliary Data (NeurIPS 2023 · similarity 0.83)
- Understanding Augmentation-based Self-Supervised Representation Learning via RKHS Approximation and Regression (ICLR 2024 · similarity 0.83)
- Adaptive Decision Boundary for Few-Shot Class-Incremental Learning (AAAI 2025 · similarity 0.82)
- Divide et Calibra: Multiclass Local Calibration via Vector Quantization (NeurIPS 2026 · similarity 0.82)
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.
