M.Sc. Chang Rajani defends his doctoral thesis "From Optics to Robotics: Machine Learning in the Real World" on Friday the 19th of May 2023 at 12 o'clock in the University of Helsinki Exactum building, Auditorium B123 (Pietari Kalmin katu 5, 1st floor). His opponent is Assistant Professor Joni Pajarinen (Aalto University) and custos Associate Professor Arto Klami (University of Helsinki). The defence will be held in English.
The thesis of Chang Rajani 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. His supervisor has been Associate Professor Arto Klami (University of Helsinki).
From Optics to Robotics: Machine Learning in the Real World
Machine learning lets a computer program learn a model of a real phenomenon from data, instead of having a person manually define rules about how the world works. Such methods are especially useful when the data is complicated, such as in the form of images or sound, since it is difficult to write down rules for them. A sub-field of machine learning called reinforcement learning further lets an agent learn to make decisions through interaction with an environment -- a trial-and-error approach to sequential decision-making.
This thesis focuses on solving difficult real-world tasks where a rule-based system would struggle, with the means of machine learning. The first task is in the field of hyperspectral imaging, a way of producing images that offer information about the scene for a continuous set of wavelengths. In contrast to the standard RGB image, spectral information lets us determine more precisely, for example, the materials which each object in a scene consists of. While spectral imaging is usually done with specialized, expensive hardware, our algorithms produce spectral information by using a standard digital camera adapted with a small passive hardware component. Our experiments show that we can produce spectral images of high quality and resolution.
The second problem we tackle is the control of an adaptive optics system. When imaging planets in other solar systems using telescopes on the ground, starlight passing through the atmosphere is distorted by turbulence. Adaptive optics uses a mirror that can be controlled to counteract the turbulence but needs an intelligent predictive algorithm running at speeds of thousands of times per second to function well. We propose the first reinforcement learning algorithms to learn a control model via interaction and demonstrate their effectiveness in simulation and on real hardware.
Finally, we tackle the problem of imitating a human demonstrator using a robot, which is a challenging transfer learning problem because humans and robots are radically different in their physical characteristics. We propose an algorithm that adapts not only the behaviour of the imitator robot but also its physical design to best imitate the human motion capture recordings.
Availability of the dissertation
An electronic version of the doctoral dissertation will be available on the e-thesis site of the University of Helsinki at http://urn.fi/URN:ISBN:978-951-51-9233-2.
Printed copies will be available on request from Chang Rajani: chang.rajani@helsinki.fi.