Image feature learning with a genetic programming autoencoder
Stefano Ruberto, Valerio Terragni, Jason H. Moore · 2020
Learning features from raw data is an important topic in machine learning. This paper presents a novel GP approach to learn high-level features from 2D images. It is a generative approach that resembles the concept of an autoencoder. Our approach executes multiple GP runs, each run generates a (partial) model that focuses on a particular high-level feature of the training images. Then, it combines the models generated by each run into a parametric function that reconstructs the observed images. We evaluated our approach on the popular MNIST dataset of 2D images representing handwritten digits. Our evaluation results show that our parametric approach can precisely reconstruct the MNIST hand-written digits.