In our lab we develop and apply state-of-the-art machine learning methods to solve a variety of problems.
Adversarial Machine Learning
Deep learning models have become a common feature in industry as well as everyday life. They are used to make decisions in autonomic driving, healthcare, security, and more. However, it has been shown that these models are vulnerable to small adversarial perturbations to their input, which can completely change their prediction. This limits the deployment of deep learning models in industrial applications. Our research concerns the analysis of these vulnerabilities and approaches for avoiding them.
Mixed-Type and Irregularly Sampled Time Series Analysis
In many real-world problems time series naturally appear to be mixed type (e.g. sensor data and event data) and irregularly sampled (e.g. clinical observations). Event data could potentially encode important context information for the interpretation of raw sensor signals like operational conditions of a system (e.g. human or device). On the other hand, more frequent hospital visits reflected in densely sampled entries in an electronic health record (EHR) could decode valuable information about the patients’ health state. Similarly, could an increasing density of error messages in an event log decode important information about the degradation of a device. Therefore, we aim develop methods which can effectively model and exploit mixed type and irregularly sampled time series for various applications.
Deep Learning for Online Handwriting Recognition
Different deep learning models have been developed for OHWR. They are used in daily life applications in mobile phones and different hand help devices. Yet, the developed models rely on specific writing tools and writing surfaces. We develop a system for OHWR using a regular ballpoint-pen equipped with several IMU sensors. Our methods rely on the data transmitted by the sensors, and allow OHWR in real time while writing on regular paper, with no use to a specific writing surface.
Deep Learning for Localization in Lidar Environments
Lidar images contain precise measurements for positions of objects, compared to RGB images. However, object detection and recognition on Lidar images are still mostly driven by conventional algorithms, compared to the current state-of-the-art in image processing: Deep Learning. One problem is the lower quantity of Lidar images, the other is the defintion and computation of locality. Our methods rely on a mixture of conventional and deep learning models to locate objects in a Lidar scene e.g. Twistlocks in Terminal Environments. Conventional models are applied to ensure robust behaviour, whereas Deep Learning models are only used to estimate a probability of object locations. Combining both methods ensures a robust and precise system for operation in terminal environment.
For many applications in various fields it is important to analyse images or videos. With the help of machine learning approaches it is possible to localise and classify objects appearing in images. Deep convolutional networks (e.g. YOLO) provide an end-to-end object detection. A very popular and widely-used application of object detection is face detection.
In our lab we use object recognition mainly in the research areas of industry and sports analytics.
Deep learning-based event detection in team sports
In team sports, the coach as well as individual athletes are heavily interested in automated recognition and analysis of on-pitch sport specific activities. In recent years, inertial measurement units in combination with machine learning methods are explored as alternative to camera-based systems. Still, the analysis of complex movement patterns remains challenging. As the amount of available data is increasing rapidly, novel deep learning methods showing promising results in related fields are applied to extend existing methods.
Clinical decision support systems
Clinical decisions may be based on subjective ratings and patient reported outcome measures which do not necessarily fully capture functional limitations, before or after an intervention. We aim to develop tools to complement those ratings by objective tools and by identifying responders / non-responders in order to provide the best treatment for the individual patient.
Federated Machine Learning for Patient-Centered Electronic Health Records
The traditional procedures for acquiring big data in machine learning (ML) involve several parties collecting the data, transferring it to a central data repository, and fusing it to build a model, whereas the data owners may be unclear about these procedures and the model future use cases. For that reason, these data acquisition procedures may violate laws such as GDPR. To address this challenge, federated learning (FL) approaches can be leveraged to build ML models such that instead of the traditional ML data transaction procedures, the model can be sent to train locally (where the data is located) and only the model updates are returned to the central system. Considering the architecture of the emerging patient-centered electronic records that store data locally under the sovereignty of individual device owners, we aim to investigate and design novel FL frameworks that enable local and distributed systems to collaboratively train a ML model without needing patients to share their raw medical information to a central entity. Moreover, to protect the data from a curious model during training, we aim to use countermeasures such as differential privacy and multi-party computation to ensure privacy guarantees.