UCLA

This doctoral dissertation is a comprehensive study on a novel method based on unitary synaptic weights to construct intrinsically stable neural systems. By eliminating the need to normalize neural activations, unitary neural networks deliver faster inference speeds and smaller model sizes while maintaining competitive accuracies for image recognition. In addition, unitary networks are drastically more robust against adversarial attacks in natural language processing systems because unitary weights are resilient to small input perturbations. The last portion focuses on a small demo that implements unitary neural nets in quantum computing. With the comprehensive performance evaluation in classical machine learning, the rigorous framework in mathematics, and the exploration of quantum computing, this dissertation establishes a solid foundation for unitary neural networks in the future of deep learning.