Projects
With AI
Tensor-based blind source separation of functional ultrasound (fUS) data
fUS is an emerging technique that can image the whole depth of the brain with unprecedented spatial and temporal resolution. The goal of this PhD research is to develop tensor-based source separation techniques that can handle the resulting large and noisy dataset and extract the activity of interest. These techniques will help to understand brain function in such great detail as never before and answer pressing neuroscience questions.
Epileptic seizure detection using tensor-based supervised learning
Epilepsy is a chronic condition that seriously compromises quality of life. Epilepsy patients could benefit from monitoring, seizure detection and prediction during daily life. Therefore, the goal of this PhD research is to facilitate supervised learning from ultralong-term multimodal data, including (wearable) data such as EEG, ECG, or motion sensors. We aim to solve specific subproblems within epileptic seizure detection/prediction, including probabilistic forecasting that can quantify the uncertainty of a prediction.
Prostate cancer detection using ultrasound
PCaVision is a novel ultrasound based alternative to diagnose prostate cancer, developed in TU/e during the last 10 years. In this collaboration, we aim to expand the existing algorithms such that they exploit the 3D nature of the images using tensor-based signal processing and machine learning techniques. Ultimately, we aim to enhance clinical performance and replace the need for invasive biopsy or costly MRI.
Identifying electropathology in atrial fibrillation
Cardiac arrhythmias are the cardiovascular epidemic of the 21th century. Unfortunately, effectiveness of arrhythmia therapy is low due to insufficient knowledge of underlying electropathology. The goal of this project is to apply big data analytics to high-resolution cardiac signals as well as surface ECG obtained during normal heart rhythms and cardiac arrhythmias in order to better understand the electropathology underlying cardiac arrhythmia.
In AI
- In this PhD project we are developing new tensor-based kernel methods that rely specifically on product kernels to enable the methods to work very efficiently for both high-dimensional and large-scale data.
- We are developing tensor-based algorithms for learning parametric models efficiently from noisy data. These models are then able to accurately predict the future. A probabilistic component will be added to the tensor-models to be able to account for noise in the measurements and make predictions with a measure of certainty.
- When it comes to the design of Bayesian machine learning algorithms, Gaussian processes (GPs) are a popular method of choice. While being flexible function approximators, GPs are capable of using the information in data sets to learn rich representations and complex structures. Unfortunately, these appealing features come at a cost of poor scalability, making GPs prohibitively expensive and unsustainable for big data. We are researching how tensor methods can help solving a large-scale GP problem in a computationally feasible way.