Neuroevolution Trajectory Networks: revealing the past of incrementally neuroevolved CNNs

Citation

Sarti S (2023) Neuroevolution Trajectory Networks: revealing the past of incrementally neuroevolved CNNs. GECCO '23 Companion: Companion Conference on Genetic and Evolutionary Computation, Lisbon Portugal, 14.07.2023-19.07.2023. https://doi.org/10.1145/3583133.3595848

Abstract
Analysing Neuroevoution algorithms often proves to be challenging from a fitness performance standpoint. We argue that our Neuroevolution Trajectory Networks (NTNs) visualisation technique, based on the use of complex networks, can effectively highlight the idiosyncratic differences and peculiarities of this family of evolutionary algorithms. This work deploys NTNs specifically to analyse the transfer of knowledge, from different benchmark classification datasets. Here, the knowledge transferred, is considered as the inheritance of evolutionary units representing the layers (and learning optimisers) forming the architecture of Convolutional Neural Networks (CNNs), incrementally developed and generated by Fast-Deep Evolutionary Network Structured Representation. Our approach highlights salient characteristics about this transfer learning paradigm, as well as exceptional findings, which help consolidate our understanding of Neuroevolution and Deep learning. This Hot-of-the-Press paper summarises the work awarded "Best Paper" entitled: "Under the Hood of Transfer Learning for Deep Neuroevolution" by Stefano Sarti, Nuno Lourenço, Jason Adair, Penousal Machado, Gabriela Ochoa; published in Applications of Evolutionary Computation. EvoApplications 2023. https://doi.org/10.1007/978-3-031- 30229-9_41