Research map / Label-efficient and robust learning / Zero-/few-shot and meta-learning

Provably Improving Generalization of Few-shot models with Synthetic Data

ICML 2025 · Label-efficient and robust learning › Zero-/few-shot and meta-learning · approach: Embedding adaptation for few-shot classification · 1 citations (OpenAlex)

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.

Develops a generalization bound for training few-shot classifiers on real plus synthetic data and turns it into a prototype/K-means-partition-based loss with discrepancy and robustness regularizers, applied to fine-tuning CLIP ViT-B/16 with Stable Diffusion-generated data.

Problem

Few-shot image classification where synthetic images from generators suffer from a distribution gap to real data, which can degrade a classifier trained on the mixed set; existing methods rely on heuristics without theoretical guarantees.

Approach

A model-dependent generalization bound over real and synthetic data (discrepancy between predictions of real and synthetic samples in each partition cell, plus local robustness) is minimized via K-means partitioning in prediction/feature space and a loss with discrepancy and robustness regularizers.

Key results

• Full version reaches 87.0% average accuracy over 10 datasets with 16 real shots, vs 86.3% for DataDream (dset) and 84.2% for real-only fine-tuning.
• Lightweight version (64 synthetic images per class, no generator fine-tuning) reaches 86.4% average, on par with DataDream.
• On CLIP-ResNet50, the method achieves 82.67 on FGVC Aircraft vs 81.46 for DataDream and 93.71 on Stanford Cars vs 93.30.

Measured effects reported in the paper

Surprising findings

• Generator fine-tuning unexpectedly degraded accuracy on DTD relative to methods without fine-tuning.
• In the extreme 1-shot setting the method underperforms DataDream (e.g., 25.3 vs 31.1 on FGVC Aircraft).

Limitations

• Underperforms the baseline in the extreme 1-shot setting, where the local robustness terms become small.
• The bound is O(sqrt(K)) in the number of partitions and is loose when only one real sample is available.
• Hyperparameters (lambda, lambda1, lambda2, cluster count) are tuned per dataset, and the cluster count is set heuristically (twice the number of classes, half for ImageNet).

Experimental setup

Datasets: ImageNet Caltech101 DTD EuroSAT FGVC Aircraft Oxford Pets Food101 SUN397 Stanford Cars Flowers102

Benchmarks: ImageNet Caltech101 DTD EuroSAT FGVC Aircraft Oxford Pets Food101 SUN397 Stanford Cars Flowers102

Metrics: accuracy

Baselines: Real-finetune IsSynth DISEF DataDream CaFo

Builds on: DataDream DISEF IsSynth RealFake LoRA

Closest papers by meaning

• Real-Fake: Effective Training Data Synthesis Through Distribution Matching (ICLR 2024 · similarity 0.89)
• Scaling Laws of Synthetic Images for Model Training ... for Now (CVPR 2024 · similarity 0.88)
• The Unmet Promise of Synthetic Training Images: Using Retrieved Real Images Performs Better (NeurIPS 2024 · similarity 0.87)
• ImagineFSL: Self-Supervised Pretraining Matters on Imagined Base Set for VLM-based Few-shot Learning (CVPR 2025 · similarity 0.87)
• Class Distribution Shifts in Zero-Shot Learning: Learning Robust Representations (NeurIPS 2024 · similarity 0.87)
• Maximizing the Potential of Synthetic Data: Insights from Random Matrix Theory (ICLR 2025 · similarity 0.87)