Research map / Label-efficient and robust learning / Zero-/few-shot and meta-learning
COBRA: COmBinatorial Retrieval Augmentation for Few-Shot Adaptation
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.
Shows that nearest-neighbor retrieval for few-shot adaptation is a special case of combinatorial mutual information and proposes COBRA, a FLMI-based submodular retrieval objective that selects relevant yet diverse samples from LAION-2B.
Problem
Nearest-neighbor retrieval for augmenting small labeled image classification datasets selects redundant auxiliary samples because it scores each sample independently and ignores diversity.
Approach
Formulate retrieval augmentation as combinatorial mutual information (CMI); existing similarity-based methods reduce to graph-cut CMI, and COBRA instead uses facility-location MI plus a soft class-balance term, optimized greedily with submodular guarantees.
Key results
- COBRA reaches 72.78% on 16-shot ImageNet with Tip-Adapter-F and ViT-B/16, above Sim-Score at 72.35%.
- COBRA has the best average rank across more than 90 settings of data scarcity, datasets, and seeds (Fig. 4).
- Retrieval with ViT-L/14 gives 78.48% for COBRA versus 78.15% for Sim-Score on ImageNet.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use COBRA retrieval instead of the strongest listed baseline, Sim-Score with soft balancing | Tip-Adapter-F on CLIP ViT-B/16 · ImageNet 16-shot | accuracy (%): 72.46 ± 0.03 -> 72.78 ± 0.13 | Helped |
| Use COBRA retrieval instead of Sim-Score | Tip-Adapter-F on CLIP ViT-B/32 · ImageNet 16-shot | accuracy (%): 67.50 ± 0.04 -> 67.76 ± 0.06 | Helped |
| Use COBRA retrieval instead of Sim-Score | Tip-Adapter-F on CLIP ViT-B/16 · ImageNet 16-shot | accuracy (%): 72.35 ± 0.01 -> 72.78 ± 0.13 | Helped |
| Use COBRA retrieval instead of Sim-Score | Tip-Adapter-F on CLIP ViT-L/14 · ImageNet 16-shot | accuracy (%): 78.15 ± 0.07 -> 78.48 ± 0.01 | Helped |
| Add the FLMI term to retrieval with class balancing and quality scoring (ablation) | COBRA on CLIP ViT-B/16 · ImageNet 16-shot | accuracy (%): 72.46 ± 0.03 -> 72.52 ± 0.11 | Helped |
| Add the quality-score term to COBRA with FLMI and class balancing (ablation) | COBRA on CLIP ViT-B/16 · ImageNet 16-shot | accuracy (%): 72.78 ± 0.13 -> 72.52 ± 0.11 | Hurt |
| Remove the class-balancing term from FLMI retrieval with quality scoring (ablation) | COBRA on CLIP ViT-B/16 · ImageNet 16-shot | accuracy (%): 72.60 ± 0.02 vs 72.52 ± 0.11 | Mixed |
| Use COBRA instead of Sim-Score for retrieval | few-shot adaptation on Flowers102 · Flowers102 | accuracy | No clear effect |
Surprising findings
- SDXL-generated synthetic images never outperform the best retrieval strategy, COBRA, on any task.
- Log-determinant MI underperforms most baselines, reaching 71.28% on 16-shot ImageNet, below no-retrieve-level gains.
Limitations
- Gains over Sim-Score are not statistically significant on Flowers102, where data is homogeneous.
- Relies on text-based prefiltering of LAION-2B captions, which restricts the auxiliary pool to captioned images containing class names.
Experimental setup
Datasets: LAION-2B ImageNet-1k ImageNet-V2 Flowers102 FGVC-Aircraft DTD Food101
Benchmarks: ImageNet-1k ImageNet-V2 FGVC-Aircraft Food-101 DTD Flowers-102
Metrics: accuracy average rank
Baselines: Sim-Score CLIP-score Random SDXL-Aug No Retrieve log det-MI MMR
Builds on: Tip-Adapter-F CaFo CLIP-Adapter Maximum Marginal Relevance
Closest papers by meaning
- Few-Shot Learning Through an Information Retrieval Lens (NeurIPS 2017 · similarity 0.85)
- Re-ranking for image retrieval and transductive few-shot classification (NeurIPS 2021 · similarity 0.85)
- Few-Shot Recognition via Stage-Wise Retrieval-Augmented Finetuning (CVPR 2025 · similarity 0.84)
- Improving Few-Shot Generalization by Exploring and Exploiting Auxiliary Data (NeurIPS 2023 · similarity 0.84)
- AR-RAG: Autoregressive Retrieval Augmentation for Image Generation (NeurIPS 2025 · similarity 0.83)
- NUDGE: Lightweight Non-Parametric Fine-Tuning of Embeddings for Retrieval (ICLR 2025 · similarity 0.83)
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.