POGP: Probabilistic Genetic Programming With Knowledge Learning for Few-Shot Image Classification

Xiaowen Huang, Yang Zhang, Zilu Wang, Le Yang, Hongsong Wang, Leimin Wang

Abstract: Genetic programming (GP) has achieved promising results without relying on the extraction of prior knowledge, e.g., fixed network architecture. However, most existing GP methods guide the evolutionary process by optimizing only the fitness function. In the few-shot image classification task, this single optimization approach can cause the model to converge prematurely to a local optimum, resulting in overfitting. Recently, knowledge learning-assisted evolutionary algorithms have demonstrated remarkable performance by making full use of evolutionary information to guide the evolutionary process. However, existing methods cannot be directly applied to a GP-based method. To address these limitations, we propose a novel framework called Probabilistic Genetic Programming with Knowledge Learning (POGP). Specifically, we design the Redress-Forget Mechanism of External Archiving to explore the potential of global information and to re-update external archives using a validation set, effectively utilizing evolutionary information while maintaining population diversity. Furthermore, the Probabilistic Initialization Strategy is designed to initialize the population using evolutionary information when it falls into a local optimum, aiming to enhance the population’s exploration ability. Moreover, we introduce Knowledge Learning-guide Evo-operations to improve the algorithm’s ability to develop optimal solutions by fusing local and global information to guide the population’s search direction. The experiments on benchmark datasets indicate that POGP outperforms state-of-the-art neural network-based and GP-based methods in almost all comparisons, demonstrating the effectiveness of the proposed POGP framework.

External IDs:doi:10.1109/tevc.2025.3642991