M.Sc. Krista Longi defends her doctoral thesis Gaussian Processes and Convolutional Neural Networks for Modeling Sensor Data on Friday the 2nd of September 2022 at 13 o'clock in the University of Helsinki Exactum building, Auditorium B123 (Pietari Kalmin katu 5, 1st floor). Her opponent is Assistant Professor Niklas Wahlström (Uppsala University, Sweden) and custos Associate Professor Arto Klami (University of Helsinki). The defence will be held in English.
The thesis of Krista Longi is a part of research done in the Department of Computer Science and in the Multi-Source Probabilistic Inference group at the University of Helsinki. Her supervisor has been Associate Professor Arto Klami (University of Helsinki).
Gaussian Processes and Convolutional Neural Networks for Modeling Sensor Data
Different sensors are constantly collecting information about us and our surroundings, such as pollution levels or heart rates. This results in long sequences of noisy time series observations, often also referred to as signals. This thesis develops machine learning methods for analysing such sensor data. The motivation behind the work is based on three real-world applications. In one, the goal is to improve Wi-Fi networks and recognise devices causing interference from spectral data measured by a spectrum analyser. The second one uses ultrasound signals propagated through different paths to localise objects inside closed containers, such as fouling inside of industrial pipelines. In third, the goal is to model an engine of a car and its emissions.
Machine learning builds models of complex systems based on a set of observations. We develop models that are designed for analysing time series data, and we build on existing work on two different models: convolutional neural networks (CNNs) and Gaussian processes (GPs). We show that CNNs are able to automatically recognise useful patterns both in 1D and 2D signal data, even when we use a chaotic cavity to scatter waves randomly in order to increase the acoustic aperture. We show how GPs can be used when the observations can be interpreted as integrals over some region, and how we can introduce a non-negativity constraint in such cases. We also show how Gaussian process state space models can be used to learn long- and short-term effects simultaneously by training the model with different resolutions of the data.
The amount of data in our case studies is limited as the datasets have been collected manually using a limited amount of sensors. This adds additional challenges to modeling, and we have used different approaches to cope with limited data. GPs as a model are well suited for small data as they are able to naturally model uncertainties. We also show how a dataset can be collected so that it contains as much information as possible with the limited resources available in cases where we use GPs with integral observations. CNNs in general require large datasets, but we show how we can augment labeled data with unlabeled data by taking advantage of the continuity in sensor data.
Availability of the dissertation
An electronic version of the doctoral dissertation is available on the e-thesis site of the University of Helsinki at http://urn.fi/URN:ISBN:978-951-51-8406-1.
Printed copies will be available on request from Krista Longi: email@example.com.