Convolutional Neural Networks (CNNs) are state-of-the-art algorithms for image recognition. To configure a CNN properly by hand, one requires deep domain knowledge, best practises, and trial and error. As deep learning progresses, manual configuration gets harder and automated ways such as genetic algorithms are required. We describe in-depth how CNN topologies can be evolved; to do so the "half pixel problem" that occurs during programmatic CNN creation and manipulation is established, analysed and solved. A new human-readable genome representation for topologies and a novel ancestry tree visualisation for genetic algorithms is used to deepen understanding of the algorithm. We rediscover common design patterns within the topologies found and see when and how the algorithm can recover from wrong assumptions in its initialisation logic. Regularisation and partial training is introduced, allowing speed-ups of up to 19% while maintaining result accuracy.
|Commitee:||Charlton, Martin D.|
|School:||University of Southampton (United Kingdom)|
|Source:||MAI 82/4(E), Masters Abstracts International|
|Subjects:||Artificial intelligence, Computer science|
|Keywords:||Evolutionary algorithms, Genetic algorithms, Image detection, Machine learning, Neural networks|
Copyright in each Dissertation and Thesis is retained by the author. All Rights Reserved
The supplemental file or files you are about to download were provided to ProQuest by the author as part of a
dissertation or thesis. The supplemental files are provided "AS IS" without warranty. ProQuest is not responsible for the
content, format or impact on the supplemental file(s) on our system. in some cases, the file type may be unknown or
may be a .exe file. We recommend caution as you open such files.
Copyright of the original materials contained in the supplemental file is retained by the author and your access to the
supplemental files is subject to the ProQuest Terms and Conditions of use.
Depending on the size of the file(s) you are downloading, the system may take some time to download them. Please be