Author: Rui Filipe Martins Monteiro
Master Degree Program in Data Science and Advanced Analytics at NOVA Information Management School
Computer Vision is a sub-field of Artificial Intelligence that provides a visual perception component to computers, mimicking human intelligence. One of its tasks is image classification and Convolutional Neural Networks (CNNs) have been the most implemented algorithm in the last few years, with few changes made to the fully-connected layer of those neural networks. Nonetheless, recent research has been showing their accuracy could be improved in certain cases by implementing other algorithms for the classification of high-level image features from convolutional layers. Thus, the main research question for this document is: To what extent does the substitution of the fully-connected layer in Convolutional Neural Networks for an evolutionary algorithm affect the performance of those CNN models? The proposed two-step approach in this study does the classification of high-level image features with a state-of-the-art GP-based algorithm for multiclass classification called M4GP. This is conducted using secondary data with different characteristics, to better benchmark the implementation and to carefully investigate different outcomes. Results indicate the new learning approach yielded similar performance in the dataset with a low number of output classes. However, none of the M4GP models was able to surpass the results of the fully-connected layers in terms of test accuracy. Even so, this might be an interesting route if one has a powerful computer and needs a very light classifier in terms of model size. The results help to understand in which situation it might be beneficial to perform a similar experimental setup, either in the context of a work project or concerning a novel research topic.
Computer Vision; Genetic Programming; Deep Learning; Convolutional Neural Networks