Image feature learning with genetic programming

Abstract

Learning features from raw data is an important topic in machine learning. This paper presents Genetic Program Feature Learner (GPFL), a novel generative GP feature learner for 2D images. GPFL executes multiple GP runs, each run generates a model that focuses on a particular high-level feature of the training images. Then, it combines the models generated by each run into a function that reconstructs the observed images. As a sanity check, we evaluated GPFL on the popular MNIST dataset of handwritten digits, and compared it with the convolutional neural network LeNet5. Our evaluation results show that when considering smaller training sets, GPFL achieves comparable/slightly-better classification accuracy than LeNet5. However, GPFL drastically outperforms LeNet5 when considering noisy images as test sets.