Generative Probabilistic Meta-Learning for Few-Shot Image Classification

Abstract

Meta-learning, a rapidly advancing area in computational intelligence, leverages prior knowledge from related tasks to facilitate the swift adaptation to new tasks with limited data. A critical challenge in meta-learning is the quantification of model uncertainty. In this paper, we propose a novel meta-learning method, Generative Probabilistic Meta-Learning (GPML), designed for few-shot image classification. GPML extends the Probably Approximately Correct-Bayes (PAC-Bayes) framework, initially formulated for single-task scenarios, to meta-learning across multiple tasks. This extension not only provides theoretical generalization guarantees for meta-learning but also effectively captures model uncertainty through variational parameters. To enhance the expressiveness of approximated posteriors in Bayesian inference, GPML incorporates implicit modeling, which defines probability distributions over task-specific parameters in a data-driven manner. This is achieved by designing a generative model structure that integrates task-dependent prior knowledge into the model inference process. We conduct extensive multidimensional performance evaluations on few-shot image classification tasks across various benchmarks, demonstrating that GPML outperforms existing state-of-the-art meta-learning methods. Additionally, ablation studies focusing on model components, the PAC-Bayes framework, and implicit modeling validate the performance improvements attributed to the proposed generative model structure, learning framework, and modeling approach.