A Comparison among Different Levels of Abstraction in Genetic Programming

In this paper we compare the performance of variants of Genetic Programming (GP) typically used for high dimensional machine learning problems. First we propose a taxonomy based on GP primitives that allow us to clearly differentiate GP variants found in literature; then we implement and test three GP variants in a set of image denoising tasks. Results show a clear advantage of the variant most commonly used for those kind of problems. We then compare our results with those reported in other GP works as well as those obtained by a Deep Neural Network (DNN). Comparisons suggest that GP cannot compete with deep learning unless it is embedded with expert’s knowledge of the problem domain.