Shape the Future of Artificial Intelligence

Join one of South Africa's leading machine learning research groups at the Faculty of Engineering, North-West University.

Applications for the 2027 intake are now open.

APPLY HERE

Before starting this application, please ensure you have your personal, academic and background information available, as well as PDF copies of your CV and academic transcript(s), as these will be required during the application process.

Join one of South Africa’s leading machine learning research groups

Scholarships Available

Master's Scholarship
R185 000* per year

PhD Scholarship
R205 000* per year

* The scholarship amounts shown include institutional funding from North-West University (such as NWU postgraduate bursaries). These bursaries do not constitute additional funding on top of the advertised scholarship value.

Who Should Apply?

Important Dates

Applications Open: Now!

Closing Date: 30 August 2026

Programme Starts: January 2027

Student Projects Available in 2027

Note: these topics are not static, but updated as new projects become available.

Understanding DNNs

Stress Testing Complexity Measures for Deep Neural Networks

This project will investigate whether we can measure how “complex” a neural network is in a way that helps us predict how well it will perform on unseen data. This is an important problem in deep learning: two models can both perform very well on the data they were trained on, but one may perform much better than the other on new data. We would like to understand whether existing complexity measures can help us tell these models apart.

The student will train a diverse collection of image classification models, such as ResNets, Wide-ResNets, Vision Transformers and ConvNeXts, using different training settings and hyperparameters. They will then implement and test several existing complexity measures and compare how well these measures predict model performance on unseen data. The project is well-suited to students interested in experimental machine learning, neural networks, computer vision and careful analysis. The expected outcome is a public collection of trained models and code that can help researchers better understand when existing complexity measures are useful, and when they fail.

Deep Neural Network Training as Node Competition

Recent work in MUST has described how the training process of a neural network can be modelled as a competitive process among nodes, with all trying to secure their role in the network by maximising their own importance. This has made it possible to better understand the conditions under which a node becomes a winner or loser, and what the implications of competition outcomes are for the training process. The “competing nodes” framework allows us to analyse some of the intriguing properties of deep neural networks, including their implicit bias towards modelling functions that not only fit the training data, but also smoothly model unseen data points. This study can either be conducted at the master's level (implementing suggested experiments, interpreting findings) or at the PhD level (developing the framework further, extending it to additional architectures and phenomena). This topic is well-suited to conceptual thinkers who enjoy grappling with technical questions and obtaining unexpected results. A strong mathematical background is required.

DNN Applications: Computer Vision

Can Medical Imaging Models Be Truly Robust?

This project will look at adversarial examples in medical imaging. An adversarial example is an image that has been changed very slightly, often in a way that is almost invisible to a person, but that causes an AI model to make the wrong prediction. A common way to defend against this is adversarial training, where the model is trained on these difficult examples so that it becomes harder to fool.

In medical imaging, however, this may be more complicated. Many clinical signs in chest X-rays are subtle: small changes in texture, brightness, shape, or structure can be important for diagnosis. This raises the question of whether the features a model should use may themselves be fragile under small image changes.
The student will train medical image classification models and study how they behave under adversarial attacks and adversarial training. A key question is whether robustly trained models still use the clinically relevant image regions, or whether they are pushed toward easier but less meaningful shortcuts, such as image quality, patient positioning, or anatomy outside the disease area.

The project is suitable for students interested in medical AI, computer vision, adversarial robustness and interpretability. The aim is to better understand when robustness methods are useful in medical imaging, and when they may change what the model is actually learning.

DNN Applications: Time Series

Domain adaptation for wearable metabolic cost estimation across walking constraints 

Indirect calorimetry remains the gold standard for measuring the metabolic cost of walking, but its three-minute response lag makes it unusable for the real-time tuning of prostheses and exoskeletons. Recent work in our group established a benchmark for predicting metabolic cost from wearable biomechanical signals across changes in walking constraint, and revealed a clear directional asymmetry in which a model generalises well from constrained to free walking but poorly in the reverse direction. This project asks the student to close that gap using domain adaptation. Working with a dataset in which the same people walked both freely and with a locked knee brace, the student will apply a family of techniques known as domain adaptation, which helps a model trained on one type of walking still work well on another. The aim is to test whether these techniques can restore the accuracy that is normally lost when the model is asked to make predictions for a walking condition it was not trained on. The project suits a candidate with a foundation in Python and deep learning who wants a well-scoped problem with a strong existing baseline, an honest test of how the model handles unfamiliar data, and a direct route to a publication.

Cross-population domain adaptation for energetics-aware prosthetic control

Most labelled biomechanics data comes from neurotypical participants, yet the clinical targets for assistive devices are people living with amputation, stroke, or other impairments, whose gait differs substantially from the training population. This project treats that mismatch as a domain adaptation problem and develops methods that transfer an energetics model learned from neurotypical or experimentally constrained walking to impaired gait, where labelled metabolic data is scarce or absent. The student will build on our temporal convolutional and adversarial adaptation work, extend it with interpretability methods that expose which gait phases and muscle activations drive metabolic cost, and validate transfer across more than one public dataset to investigate cross-population generalisation. A strong candidate will combine deep learning, time series modelling, and a real interest in human movement, and will graduate with a body of work positioned at the intersection of explainable machine learning, domain adaptation, and rehabilitation engineering.

The Beyond Limits Prosthetic Hand

Development of a responsive and affordable prosthetic hand is a long-term goal of the MUST/Beyond Limits collaboration. An arm bracelet, capable of capturing surface electromyography (sEMG), is used to analyse and predict the intention of the wearer. Initial models based on Long Short Term Memory (LSTM) networks have been found to be able to predict finger angle movement with high accuracy, using both time- and frequency-based preprocessing that is fast and computationally cheap. Current work focuses on estimating intended grip strength in addition to finger movement, integrating modelled components with the microcontroller on the current hand prototype, exploring different architectures, and the use of subject-specific finetuning of a larger model. As this is a rapidly moving project, the exact master’s level studies linked to it will only be defined at the start of 2027. This study is suited to students interested in empirical machine learning, solving practical problems, and making a direct impact through their work.