Generalisation in Deep Learning

A long-standing theoretical interest in our group focuses on the topic of generalisation - that is, how information learned on a training set of samples is transferred to novel inputs. Investigations on generalisation in Deep Neural Networks (DNNs) are particularly relevant in light of recent progress in the deep learning field, and is an important part of our current work. Currently, MUST students are studying the essence of the learning process of different types of deep networks, investigating how factors such as learning algorithms and network structures affect generalisation. We typically use popular deep learning tools (such as Pytorch and Tensorflow) to explore specific questions on new and existing data sets. 

NITheCS

Machine learning in support of Theoretical and Computational Sciences is a research programme of the National Institute for Theoretical and Computational Sciences (NITheCS). Collaborators include researchers from diverse fields at UCT, UKZN, UJ, SU, and NWU. 

Coordinated by MUST, the programme has two streams: 

• Machine learning research: development of new, specialised ML techniques.
• Machine learning as a tool: applying ML for scientific modelling applications.

The current focus of the programme is on knowledge discovery in time series data, and on creating a forum for cross-cutting projects executed in other focus areas where projects rely on machine learning expertise.    

NITheCS ML reseach programme

Speech and language processing

MUST collaborates with Saigen to solve real-world speech analytics problems through the application of deep learning techniques. Examples include using DNNs to improve automatic speech recognition and speaker diarisation (identification of individual speakers) on heavily compressed or poor quality audio files, and the study of DNN-based word embeddings for more accurate language modelling.

Speech processing tools are not only applicable to human speech and language: with Prof Jaco Versfeld from Stellenbosch University we are investigating the use of end-to-end DNN systems for whale call detection and classification. Audio data collected through passive acoustic monitors in False Bay sheds light on the movement of local whale species, information that supports marine monitoring and conservation. Our role is to develop tools to contribute to this process.

CAIR

Space weather prediction

Together with the Space Science directorate of the South African National Space Agency (SANSA), we investigate the applicability of DNNs for modelling various space weather phenomena, such as  predicting geomagnetic disturbances from solar wind parameters, or predicting the eruption of solar flares from images of the sun. An important goal of this work is ‘knowledge inference’: studying ways in which the developed models can be interpreted in novel ways in order to shed light on the underlying phenomena.

Industry applications of deep learning

DNNs have wide application in industry. As part of our industry programme, we investigate the ability of DNN-based tools to solve very specific tasks in close collaboration with external domain partners. In these cases, the industry partner contributes the domain knowledge, and MUST researchers tailor deep learning models to address the specific challenges posed by the task.

-  Channel state estimation

Together with Prof Albert Helberg and the NWU Telenet Group, we investigate the abilities of deep learning methods in the telecommunications domain.  Specifically, we consider the applicability of generative adversarial networks to channel estimation and equalisation over wireless channels such as WiFi. We hope to further extend the application and usefulness of deep learning methods to additional telecommunication tasks.

-  SpotD

This is a Technology Innovation Fund (TIA) funded project that is developing an automated image change detection platform for a wide range of application domains including diagnostics, integrity verification, and remote sensing. The project seeks to streamline the image comparison process, resulting in time savings and improved robustness and accuracy. By leveraging advanced image processing technologies, the platform aims to provide professionals with a user-friendly interface to analyze and compare images efficiently, using the right tools for each particular use case. The project's focus on practical applications makes it a promising venture with the potential for impact in various industries.

-  Sailplane cross-country performance optimisation

In collaboration with Jonker Sailplanes, we are developing a framework for optimising sailplane cross-country performance with DNNs. Traditional sailplane performance optimisation is a slow and computationally expensive task as it requires the integration of multiple simulation packages and function evaluations for multiple span-wise stations (parts of a wing) under multiple flight conditions. We are replacing the computationally expensive simulation packages with lower fidelity DNN-based models, in order to reduce the time spent on the preliminary sailplane design and optimisation phase. Dr Johan Bosman, an aerodynamics engineer at Jonker Sailplanes and senior lecturer at the School for Mechanical Engineering of the Faculty of Engineering, co-supervises the study. 

Completed projects

The Babel Project

The Speech Transcription Platform Project

 Google Text To Speech Project

The SADE Project

The V-BAT Project

The VOICES Project

The Lwazi Project