Research map / Image generation / Autoregressive and latent generation
Leverage Class-Specific Accuracy to Guide Data Generation for Improving Image Classification
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.
Proposes needs-aware image generation: a multi-level optimization framework that allocates a limited class-conditional GAN (BigGAN) generation budget to classes according to their validation loss/accuracy, trained end-to-end with the classifier.
Problem
Limited labeled training images cause overfitting in image classification; existing synthetic data generation allocates budget per class based on training-set frequency rather than classes' actual data needs (classification performance).
Approach
Measure per-class validation loss of an initial classifier and sample generated class labels proportionally to it, then jointly optimize classifier, class-conditional GAN, and meta parameters (shared last layers/architecture) via a four-stage multi-level optimization so generation and classification mutually influence each other.
Key results
- On CIFAR10-LT with imbalance factor 200, the method reaches 73.41% test accuracy vs 68.91% for Meta weight and 65.68% for vanilla ResNet; on CIFAR100-LT it reaches 40.95%.
- Applied on top of CC-SAM, the method improves CIFAR10-LT (imbalance 10) from 80.94 to 81.89 and CIFAR100-LT (imbalance 200) from 45.66 to 47.09.
- On balanced ImageNet with PC-DARTS, top-1 error drops from 24.2% to 23.2% and top-5 error from 7.3% to 6.6%; LiVT + Ours improves ImageNet-LT accuracy from 60.9 to 61.8.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| allocate generated images according to class-specific performance rather than class frequency | image classification · CIFAR10-LT and CIFAR100-LT | test accuracy (%) | Helped |
| apply needs-aware generation to LiVT | LiVT · ImageNet-LT, Few classes | top-1 accuracy (%): 41.0 -> 43.9 | Helped |
| apply needs-aware generation to PCDARTS | PCDARTS · ImageNet test set | top-1 classification error (%): 24.2 -> 23.2 | Helped |
| perform the framework stages end-to-end rather than sequentially (ablation) | needs-aware image generation · CIFAR10-LT, imbalance factor 200 | test accuracy (%): 69.24 -> 73.41 | Helped |
| perform the framework stages end-to-end rather than sequentially (ablation) | needs-aware image generation · CIFAR100-LT, imbalance factor 200 | test accuracy (%): 37.79 -> 40.95 | Helped |
| replace BigGAN with VCGAN (ablation) | needs-aware image generation · Imbalanced CIFAR-10, imbalance factor 200 | test accuracy (%): 73.41 -> 72.33 | Hurt |
| replace BigGAN with ACGAN (ablation) | needs-aware image generation · Imbalanced CIFAR-10, imbalance factor 200 | test accuracy (%): 73.41 -> 72.06 | Hurt |
| replace BigGAN with VCGAN (ablation) | needs-aware image generation · Imbalanced CIFAR-100, imbalance factor 200 | test accuracy (%): 40.95 -> 39.88 | Hurt |
| replace BigGAN with ACGAN (ablation) | needs-aware image generation · Imbalanced CIFAR-100, imbalance factor 200 | test accuracy (%): 40.95 -> 40.07 | Hurt |
| apply needs-aware generation to PCDARTS instead of generating an equal number of images per class | PCDARTS · ImageNet test set | top-5 classification error (%): 7.2 -> 6.6 | Helped |
| apply needs-aware generation to PCDARTS instead of training CGAN and classifier separately | PCDARTS · ImageNet test set | top-1 classification error (%): 24.1 -> 23.2 | Helped |
Surprising findings
- Classes with fewer real training examples do not necessarily have worse accuracy; e.g., horse (368 images) outperforms dog (568 images) before augmentation in CIFAR-10-LT.
- Sequential (non-end-to-end) execution of the stages performs worse than the end-to-end framework (69.24 vs 73.41 on CIFAR10-LT).
Limitations
- Cannot be applied to reinforcement learning or evolutionary NAS methods since they are not differentiable.
- Relies on a generator trained on the same small dataset; generated image quality depends on GAN capacity (BigGAN).
Experimental setup
Datasets: CIFAR10-LT CIFAR100-LT ImageNet-LT ImageNet
Benchmarks: CIFAR10-LT CIFAR100-LT ImageNet-LT ImageNet
Metrics: test accuracy top-1 error top-5 error many/medium/few accuracy human rating of naturalness and correctness inception score FID
Baselines: Reciprocal Separate MTL Equal MiSLAS GCL CC-SAM RIDE CMO LiVT PC-DARTS
Builds on: multi-level optimization BigGAN Gumbel-Softmax PC-DARTS bi-level optimization
Closest papers by meaning
- Synthetic Tabular Data Generation for Imbalanced Classification: The Surprising Effectiveness of an Overlap Class (AAAI 2025 · similarity 0.85)
- Generative Classifiers Avoid Shortcut Solutions (ICLR 2025 · similarity 0.85)
- Real-Fake: Effective Training Data Synthesis Through Distribution Matching (ICLR 2024 · similarity 0.84)
- Provably Improving Generalization of Few-shot models with Synthetic Data (ICML 2025 · similarity 0.84)
- The Unmet Promise of Synthetic Training Images: Using Retrieved Real Images Performs Better (NeurIPS 2024 · similarity 0.83)
- Salient Concept-Aware Generative Data Augmentation (NeurIPS 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.