Created by W.Langdon from gp-bibliography.bib Revision:1.8098
We begin the thesis using a genetic algorithm to optimize a U-Net's architecture and compare our results with the original U-Net model. We show that the genetic algorithm can improve the performance of the original U-Net. Next, we refine our method by adding more hyperparameters and different architectures into the evolutionary optimisation process. The genetic algorithm was able to find a smaller model than the original U-Net with only a small trade-off in the area under the curve (AUC) metric observed. We next examine different fitness functions to improve computational performance to reduce overfitting. We improved our previous work and also showed that genetic algorithm could find better performance models compared to the manual design ones.
We next use the genetic algorithm to optimize the Spatial Attention UNet, which has been shown to be less prone to over-fitting as it puts more emphasis on more important features in the images. The results show smaller models with similar AUC compared to state-of-the-art models. We also improve training time by using custom epoch optimization and reducing duplicated individuals. This reduced training time by more than half, enabling us to use that saved time to find a better model than in previous works.
Finally, we change strategy and use grammatical evolution (GE) instead of the genetic algorithm as our evolutionary strategy. We achieve better AUC than the genetic algorithm and evolve a 10 times smaller model with similar accuracy.
This thesis shows that evolutionary algorithms successfully improve state-of-the-art neural network models in image segmentation.",
Supervisor: Conor Ryan and Patrick Healy",
Genetic Programming entries for Mahsa Mahdinejad