Under the Hood of Transfer Learning for Deep Neuroevolution

Citation

Sarti S, Laurenço N, Adair J, Machado P & Ochoa G (2023) Under the Hood of Transfer Learning for Deep Neuroevolution. In: Applications of Evolutionary Computation. EvoApplications 2023., 12.04.2023-14.04.2023. Springer Nature Switzerland, pp. 640-655. https://doi.org/10.1007/978-3-031-30229-9_41

Abstract
Deep-neuroevolution is the optimisation of deep neural architectures using evolutionary computation. Amongst these techniques, Fast-Deep Evolutionary Network Structured Representation (Fast-DENSER) has achieved considerable success in the development of Convolutional Neural Networks (CNNs) for image classification. In this study, variants of this algorithm are seen through the lens of Neuroevolution Trajectory Networks (NTNs), which use complex network modelling and visualisation to uncover intrinsic characteristics. We examine how evolution uses previously acquired knowledge on some datasets to inform the search for new domains with a specific focus on the architecture of CNNs. Results show that the transfer learning paradigm works as intended as networks mutate, incorporating layers from the best models trained on previous datasets. The use of specifically designed NTNs in this analysis enabled us to perceive the architectural characteristics that evolution favours in the design of CNNs. These findings provide novel insights that may inform the future creation of Deep Neural Networks (DNNs).

People

Dr Jason Adair

Lecturer in Data Science, Computing Science

Professor Gabriela Ochoa

Professor, Computing Science

Mr Stefano Sarti

Tutor, Computing Science and Mathematics - Division