Unresolved problems in deep neural networks: Causes and solutions

The aim of the proposed research is the identification of the causes of problems that limit the practical use of deep neural networks because they may return incorrect answers (outputs), and the development and implementation of methods that solve these problems.

The proposed research will increase the practical use of deep neural networks, especially in life-critical applications, for example, medical diagnosis and computer vision systems in self-driving cars because an incorrect output from a deep neural network may have serious consequences.

The research involves theoretical development using advanced mathematics, computational implementation and then testing on benchmark datasets that are in the public domain.

There are many issues to be considered when designing a deep neural network, and the research will be guided by general principles, rather than a restriction to a specific type of deep neural network. It is hoped that this generalisation will increase the applications of the research.