Seminar - Evolutionary Deep Learning for Image Classification

ECS PhD Proposal

Speaker: Bin Wang
Time: Monday 21st October 2019 at 09:30 AM - 10:30 AM
Location: Cotton Building CO431

Add to Calendar Add to your calendar

Abstract

Along with the boosted demand for image classification, researchers have manually designed deep convolutional neural networks (CNNs) to improve classification accuracy with domain knowledge, which is time-consuming and error prone. Therefore, automatically designing CNNs has drawn interests, which, however, suffers from the intimidating computation cost. In addition, compact CNNs have emerged to reduce the size without compromising the accuracy, but there is limited work. To address the above limitations, this research will propose evolutionary computation methods to automatically evolve optimal CNNs with four major contributions. Firstly, this research will develop new fitness evaluation methods to significantly reduce the computational cost. Secondly, a coevolution algorithm will be designed and developed to simultaneously evolve the macro-architecture and the micro-architecture to better explore the CNN architecture space. Thirdly, a multi-objective co-evolution algorithm will be designed and developed to evolve a Pareto-front of trade-off, so the end-user can choose the most suitable solution according to the computational resource and the requirement of classification accuracy. Lastly, a visualisation tool will be developed to analyse the evolutionary process to further improve performance.

Go backGo back to the seminar list