Research map / Label-efficient and robust learning / Zero-/few-shot and meta-learning
Regularizing Neural Networks with Meta-Learning Generative Models
Shin'ya Yamaguchi, Daiki Chijiwa, Sekitoshi Kanai, Atsutoshi Kumagai, Hisashi Kashima
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 cross-entropy on synthetic samples with pseudo consistency regularization (PCR) (ablation) | GDA · Cars | top-1 accuracy (%): 84.50 -> 86.36 | Improved |
| add meta pseudo sampling (MPS) to PCR (ablation) | PCR / MGR · Cars | top-1 accuracy (%): 86.32 -> 87.22 | Improved |
| add MPS to plain GDA (ablation) | GDA · Cars | top-1 accuracy (%): 84.50 -> 85.28 | Improved |
| optimize generator G_Phi instead of the finder F_Phi (ablation) | MPS · Cars | top-1 accuracy (%): 87.22 -> 84.47 | Hurt |
| use MLP finder instead of residual tanh finder (ablation) | MPS finder · Cars | top-1 accuracy (%): 87.22 -> 86.35 | Hurt |
| add KL penalty on finder outputs (ablation) | MPS · Cars | top-1 accuracy (%): 86.92 -> 87.22 | Improved |
| use 1st-order finite-difference approximation instead of 2nd-order gradients (ablation) | MPS · Cars | top-1 accuracy (%): 87.30 -> 87.22; 6.55 -> 5.79 hours | No clear effect |
| use a better generator (FastGAN to StyleGAN-XL) (ablation) | MGR · Cars | top-1 accuracy (%): 86.30 -> 88.37 | Improved |
| use synthetic samples instead of real samples for consistency regularization (ablation) | MGR · Cars | top-1 accuracy (%): 86.16 -> 87.22 | Improved |
| combine MGR with TrivialAugment | TrivialAugment base · Cars | top-1 accuracy (%): 87.83 -> 89.10 | Improved |
| train finder by maximizing CE loss instead of meta-learning (ablation) | MPS · Cars | top-1 accuracy (%): 87.22 -> 86.52 | No clear effect |
| add LatentAugment on top of RandAugment (ablation) | MGR · Cars | top-1 accuracy (%): 87.22 -> 87.85 | Improved |
Closest papers by meaning
- Do We Need All the Synthetic Data? Targeted Image Augmentation via Diffusion Models (ICLR 2026 · similarity 0.88)
- Deep Generative Modeling on Limited Data with Regularization by Nontransferable Pre-trained Models (ICLR 2023 · similarity 0.87)
- Toward Understanding Generative Data Augmentation (NeurIPS 2023 · similarity 0.87)
- A Bayesian Data Augmentation Approach for Learning Deep Models (NeurIPS 2017 · similarity 0.87)
- MGDD: A Meta Generator for Fast Dataset Distillation (NeurIPS 2023 · similarity 0.86)
- Data Augmentation with Diffusion for Open-Set Semi-Supervised Learning (NeurIPS 2024 · similarity 0.85)
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.
