Posted by annar on Fri 01 Nov 2019, 11:15:59 Fri 01 Nov 2019, 11:15:59

Dr Anna Bosman received her PhD degree in September this year with a thesis titled "Fitness Landscape Analysis of Feed Forward Neural Networks". Neural network training is a highly non-convex optimisation problem with poorly understood properties. Due to the inherent high dimensionality, neural network search spaces cannot be intuitively visualised, thus other means to establish search space properties have to be employed. In her thesis, Dr Anna Bosman applied fitness landscape analysis techniques to gain insight into the nature of the neural network search spaces. A novel gradient-based sampling technique was proposed, together with a novel method to quantify and visualise the modality of neural network error landscapes. The proposed method creates 2-dimensional projections of complex multidimensional search spaces, and enables empirical and visual investigation thereof. Using the proposed method, empirical evidence for existing theoretical conjectures was obtained, thus contributing to the fundamental understanding of neural networks. Research conducted for the PhD was presented at multiple international conferences, and published in the Neural Processing Letters journal.