Improving neural networks by reduction of parameters and noise injection

Thesis

Deep learning has been revolutionising the way we look at prediction, signal processing, perception, classification, RL, control and many others. Neural networks have become the centre of attention of the scientific community re- volving around computational solutions to difficult problems.

While in principle such models can be used in their pristine form to approxi- mate complex functions, it is important to point out rather high requirements during training and evaluation.

This thesis is concerned with directly supervised problems - a setting funda- mental to the core deep learning. It is the most common incarnation of the neural revolution.

Improving supervised training of deep networks can have a critical impact on the future of prediction based on labels, which is becoming the tool of choice for solving a rapidly growing number of challenges.

Authors: Moczulski, M

• DOI: 10.5287/ora-av5rk8ar8
• Type of award: DPhil
• Level of award: Doctoral
• Awarding institution: University of Oxford
• Language: English
• Keywords: supervised learning; neural networks; deep learning
• Subjects: Deep learning (Machine learning); Neural networks (Computer science); Supervised learning (Machine learning)