Transforming GP-CNN Tree Search Into Trainable Architectures for Image Classification

Data-efficient image classification poses a challenge in achieving effectiveness with limited data, as evidenced by the current methods based on convolutional neural networks (CNNs) and genetic programming (GP). Existing works employing these two methods encounter limitations, such as a lack of flexibility and an inability to effectively explore the latent features of the data. To tackle these challenges, this paper introduces a genetic programming method for data-efficient image recognition, leveraging novel function sets, terminal sets, and program structures. This method transforms tree-based data structures in GP into trainable CNN architectures. Further, by employing block structures instead of single operations in the search space, the search space is reduced and the stability of the search structures enhanced. Comparative experiments with state-of-the-art neural network methods and GP-based methods on data-efficient classification datasets validate the GP-CNN method offering higher performance.