Dave Marshall's PhD Availability Page

I am basically interested in supervising PhDs in the areas of Computer Vision, Digital Audio, Computer Music and Machine Learning.

Please email me for more details or to discuss other possible PhD projects.

School Scholarships are available --- details here.

Self-funded or externally funded applications welcome.

Some Sample Available Projects

Seeing the Future: Predicting Video

The estimated rate that information enters the human eye is a staggering 72GB/s. Using very sluggish neurons and a power budget of a few watts the human brain processes this information, creates a percept of a stable world and detects important events occurring within the world, significantly outperforming conventional computer vision methods. How does the brain achieve such efficient and fast processing? Predictive coding is a term used by brain scientists to describe an idea that has its roots in the earliest days of psychology, and had a later echo in the field of cybernetics. The theory proposes that in order to drastically reduce the inherent data processing requirements, and thus achieve efficiency, the brain tries to predict incoming sensory information. It does this for two reasons. First, the processing demands for testing a prediction of what something is are considerably lower than for deducing what something is. Second, prediction failures indicate sensory input that merits extra processing resources, where something is changing or incongruent.

We have recently completed an EPSRC funded project to build the first artificial predictive vision system, which we now wish to build upon. The potential applications of this technology are broad (e.g. healthcare, security, ubiquitous computing), and the technology could be used as part of a standalone computer-based system, or to help or augment a human operator.

In order to predict scenes, an understanding of the scene captured in the current and previous frames is required. For this purpose, we propose to utilise off-the-shelf publicly available research code (in-house, standard toolboxes etc.). For example, methods that can analyse basic motion in the scene (optical) and from this high level physics based motion (e.g. moving car) or human activity (e.g. walking pedestrian) are required. The student will evaluate current available methods on our newly acquired data set.

A variety of PhD projects set around this area of research are available:

Neural models for prediction of brain activity from fMRI and MEG data sources.

The brain can essentially be regarded as a complex neural network.
The brain has remarkable predictive capabilities.
Inspired by models of how the brain works we are interested in for prediction of brain activity from fMRI and MEG data sources.

Independent Component Analysis (ICA) is regularly applied to fMRI and/or MEG data to identify and segment sub-networks and regions. In signal processing, independent component analysis (ICA) is a computational method for separating a multivariate signal into additive subcomponents. This is done by assuming that the subcomponents are non-Gaussian signals and that they are statistically independent from each other. ICA is a special case of blind source separation and the classic illustration of the technique is in solving the "cocktail party problem" – many speakers small number of input sources (microphones) decouple the sounds so you can hear individuals (Humans can do this well).

We plan to use ICA (and similar associated techniques such as Principal Component Analysis or PCA) to model brain activity that is recorded from fMRI and MEG sources.

We will evaluate (and subsequently refine) the model by training on one dataset and then testing on an independent dataset.

Once we have built a satisfactory model we will use it to predict neural activity in one location in the brain given known activation in other brain regions. We will use Machine Learning, Deep learning and time series data analysis for this stage.

When the input or tasks demands are changed, discrepancies between predicted activity and measured activity will be used to identify regions and networks involved in processing specific stimuli, for example faces or text. As a second stage we will look to extend the technique to work across modalities – to visual input or between fMRI and MEG.

Computational Music and Machine Learning

Prof. David Marshall and Dr. Kirill Sidorov established the Computational Music research within the School a few years ago following their innovative paper on “Music Analysis As a Smallest Grammar Problem”, ISMIR 2014 (http://www.terasoft.com.tw/conf/ismir2014/). We can successfully analyse the high-level musical structure. Furthermore, we can edit this structure to produce new music similar to the original and wide range of other possible applications exists, including automatic summarization and simplification; estimation of musical complexity and similarity, and plagiarism detection.

We are interested in many Music related problems that involve computational analysis especially using Machine Learning, Deep Learning etc. These could involve audio, computer music representations, imagery (Sheet music) or video, or mixtures of any of these.

Some possible example projects include (but are not exclusive):

The above are merely suggestions of possible PhD projects and variations on the themes of the above and, indeed, any related are ideas. Please get in touch: MarshallAD@Cardiff.ac.uk to discuss any such potential PhD projects.