Seminar - Evolutionary Computation for Interpretable Tree-like Artificial Neural Networks
ECS PhD Proposal
Speaker: Damien O'Neill
Time:
Friday 22nd March 2019 at 09:00 AM -
10:00 AM
Location:
Cotton Club,
Cotton 350
Abstract
Convolutional neural networks(CNNs)are considered a state-of-the-art machine learning technique for image classification, however the space of possible CNN architectures remains relatively unexplored. In particular, tree-like CNNs are an unexplored CNN architecture which may enhance the efficiency and interpretability of CNNs. Neuroevolution offers a possible solution to the problem of CNN architecture exploration, including tree-like CNNs, by providing an automatic search through the space of possible CNN architectures, but current neuroevolution algorithms have not been formulated to search for tree-like CNN architectures. The goal of the current project is to formalise and examine important characteristics of tree-like CNNs, and develop an efficient novel tree-based genetic programming neuroevolution algorithm to search for high performance tree-like CNN architectures. Further, the current work will develop a novel evolutionary multi-objective algorithm to search for a set of trade-off solutions in terms of characteristics of tree-like CNNs, and will also develop a novel visualisation technique to aid in understanding inference in CNNs, with an emphasis on understanding inference in tree-like CNNs.