University of Twente Student Theses


Extracting Unseen Classes From Capsule Network Feature Space

Couwenberg, F. (2022) Extracting Unseen Classes From Capsule Network Feature Space.

[img] PDF
Abstract:A shortage of healthcare professionals already is the norm, and the shortage will continue to increase even more in the next years. Autonomous robotics might be able to provide some relief for these professionals and our healthcare system. However, robots have not yet been able to achieve enough autonomy to function in such a setting with many unknowns. Incremental learning would allow the robot to learn on the job and therefore adapt to new situations. In this work, I argue that incremental learning can best be achieved by learning a broad basis of elements and combining these into new classes. In this research, two capsule networks are built and investigated: Efficient CapsNet and Small CapsNet. The results indicate that Small CapsNet is able to learn generalizing factors, but cannot yet generalize to data augmentations it has not seen before. In addition, both Efficient and Small CapsNet show that it is possible to learn a new class, based only on the information already known. The results indicate that a broad basis of classes is necessary, but that as long as the unfamiliar class contains similar elements it should be possible to combine these separate elements into a new class.
Item Type:Essay (Master)
Faculty:ET: Engineering Technology
Subject:50 technical science in general, 54 computer science
Programme:Biomedical Engineering MSc (66226)
Link to this item:
Export this item as:BibTeX
HTML Citation
Reference Manager


Repository Staff Only: item control page