Information Component Analysis via Deep Learning

Project description

As their prominent characteristics, perceptual data often convey the mixing information, which often results in the inadequate performance for a specific perceptual information processing task due to the interference of irrelevant information components. For example, facial images typically convey the mixing information including identity and expressions. For specific tasks like face and facial expression recognition, the mixing information components are hardly separable, which results in difficulties in either of two tasks. The same problem also exists in speech information processing where speech conveys the mixing information including linguistic, speaker, emotional and environmental characteristics. Furthermore, there is no equal amount of information for mixing components; e.g. linguistics often overwhelmingly dominates the information in speech. The nature of perceptual data gives rise to considerable challenges in their modelling, analysis and recognition.

The project is going to investigate and develop a generic approach to information component analysis for perceptual data with state-of-the-art machine learning techniques, deep learning. Surrounding the main theme on how to disentangling/extracting information components, main issues to be studied include objective-driven high level abstraction of perceptual data in flexible representation forms, novel building blocks and deep learning models including architectures and learning algorithms to carry out an information component "filter' and theoretic information aspects in measuring the extracted information components. For demonstration, an information component analysis prototype would be developed for a real application, e.g., speech or facial information component analysis. In general, this project is suitable for one who is interested in fundamental research in machine learning while it is acceptable for one who has a relevant application problem in mind and wishes to tackle their problems with an emerging technology such as deep learning.

It is worth highlighting that the hypotheses set in this project are original and hence this is an extremely challenging project of a great novelty. In order take this project, thus, it is essential to be highly self-motivated and to have excellent background knowledge in mathematics, machine learning, image or speech signal processing and good programming skills. If you are interested in this project, please first visit my research student page: http://staff.cs.manchester.ac.uk/~kechen/ for the required materials and information prior to contacting me.