Generalisability of U-Net models evolved using grammatical neuroevolution

In deep learning, the performance of a model for different datasets depends heavily on its architecture and choice of hyper-parameters. Choosing these appropriately is a difficult task, some will work well with some datasets, while they may perform poorly in others. The process of optimizing the architecture and collection of hyper-parameters, in order to achieve high accuracy and reduce overfitting, using evolutionary algorithms is known as neuroevolution. Various optimization techniques have been used for neuroevolution, in this paper, we consider convolutional neural networks optimised using grammatical evolution. We have already shown that grammatical evolution can successfully tune U-Net hyper-parameters and boost its performance in iris blood vessel segmentation (DRIVE dataset). However, we wish to examine the generalisation of the evolved models. To show this, we test the best models found on the DRIVE dataset on one other iris blood vessel segmentation benchmark and in a vastly different domain, flood-affected area detection. Our results show that the models found during evolution exhibit good performance on other datasets and that transfer learning in grammatical evolution based neuroevolution is a fruitful avenue for future research.