Note (General)Deep neural networks (DNNs) such as convolutional neural networks (CNNs) have enabled remarkable progress in the application of machine learning and artificial intelligence. Research scientists are gearing up for adopting DNN methods to their respective domain problems. Automated neural architecture search (NAS), also known as automated DNN (AutoDNN), aims to automate the architecture search of neural networks to enable researchers adopt DNN methods with ease, and with little or no expertise in deep learning. As metaheuristic approach, automated NAS requires a representation scheme to encode the candidate solutions (architectures). Direct encodings of genetic algorithms and genetic programming have been widely employed in automated NAS methods. Though easy to implement, direct encoding cannot be easily modularized and the lack of distinctive separation of genotype and phenotype spaces limits their functional complexity. Therefore, it may be difficult for direct encodings to evolve modules (building-blocks) with shortcut and multi-branch connections which can improve training and enhance network performance in image understanding tasks. This work presents a novel generative encoding, called symbolic linear generative encoding (SLGE), that combines the complementary strengths of gene expression programming (GEP) and cellular encoding (CE) for automatic architecture search of deep neural networks for image understanding. In particular, evolving modularized CNNs with shortcut and multi-branch modularity properties (similar to the ones commonly adopted by human experts) for remote sensing (RS) image understanding tasks such as scene classification and semantic segmentation. GEP is known for its simplicity in implementation and multi-gene chromosomes with flexible genetic modification, whereas CE has the ability to produce modular artificial neural networks (ANNs). Both GEP and CE are well established evolutionary computation methods which have experienced a lot of development and theoretical study. A large part of this previous work involves architecture search of ANNs in a small scale, and therefore this work provides the possibility for CNNs architecture development for image understanding tasks, particularly in the field of RS. We adopt two automated NAS search strategies: random search with early-stopping and evolutionary algorithm, to automatically evolve modularized CNNs architectures for classification of RS imagery scenes and semantic segmentation of aerial/satellite imagery respectively. Two types of multi-class image scene classification tasks were performed: single-label scene classification and multi-label scene classification, using four different remotely-sensed imagery datasets, to validate the expressiveness and tractability of SLGE representation space. Moreover, we constructed a two-separate SLGE representation spaces: normal cell and atrous spatial pyramid pooling (ASPP) cell. Then, using evolutionary algorithm with genetic operators such as uniform mutation, two-point crossover and gene crossover, we joint search for a normal cell and an ASPP cell as a pair of cells to build a modularized encoder-decoder CNN architecture for solving RS image semantic segmentation problem. Three RS semantic segmentation benchmarks were used to verify the performance of the SLGE architecture representation. By doing this, we also validated the effectiveness and robustness the proposed SLGE architecture representation. The results position SLGE architecture representation amongst the best of the state-of-the-art systems.
Collection (particular)国立国会図書館デジタルコレクション > デジタル化資料 > 博士論文
Date Accepted (W3CDTF)2023-10-11T15:41:03+09:00
Data Provider (Database)国立国会図書館 : 国立国会図書館デジタルコレクション