Diversity-Preserving Few-Shot GAN Adaptation via Layer-Aware and Similarity-Guided Meta-Training Samples

Résumé fourni par la source

Few-shot learning is a significant challenge for Meta-GANs. Without explicit mechanisms to preserve diversity, the model tends to overfit to the scarce training data. We build on a Reptile-based meta-learning backbone and extend it into a diversity-preserving Meta-GAN framework. Our model aims to improve both the stability and diversity of few-shot adaptation. Although the meta-training phase follows the standard Reptile mechanism, the meta-testing phase introduces several key modifications to address the diversity collapse that is typical of low-data adaptation. We employ a strategy in which one copy of the meta-trained model is reused to induce diversity throughout the meta-test training, whereas another copy is adapted to the target task. To ensure fidelity to the target sample, we use layer freezing and train exclusively on the target sample and meta-training samples during alternating discriminator iterations. Compared to baseline Reptile-MetaGAN experiments on the MNIST and Omniglot datasets in a few-shot configuration yielded improvements in Fréchet Inception Distance (FID), Kernel Inception Distance (KID), and Inception Score (IS). These findings demonstrate that a stable and diversity-aware pathway for few-shot generative adaptation can be achieved using cross-task diversity meta-train samples, alternating layer freezing, and selective SSIM filtering.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.