Friday, 24 November, at 13:00, The Technical University of Denmark, Matematiktorvet, Building 303A, Auditorium 45.
Summary:
The statistically derived data models called Convolutional Neural Networks have revolutionized the field of Computer Vision. Their ability to interpret images and recognize faces, categorize objects, detect location of objects in video etc. has led to a vast recent use of them.
The problem with the networks is their need of large quantities of well-structured and labeled data in order to perform well. Gathering these large datasets can be a challenge and sometimes infeasible, which largely limits their use.
In this thesis methods to understand these models better and tools to learn from them despite lack of sufficiently large datasets are provided. These methods and tools are both new ways of testing the networks ability to generalize but we also show how simulated data can play a big role in stabilizing the models.
We can conclude that there is a large potential in learning from Convolutional Neural Networks even on smaller datasets if proper tests and methods are used to understand their functionality. The results of this thesis provide a foundation for future research in data-driven modeling to learn about the limitations and challenges of the available data.
Read more about this thesis in DTU Orbit.
Supervisor: Associate Professor Allan Aasbjerg Nielsen, DTU Compute.
Co-supervisor: Associate Professor Henning Skriver, DTU Space.
Co-supervisor: Senior Specialist, Rasmus Engholm, Terma A/S.
Co-supervisor: Senior Systems Engineer, Morten Østergaard Pedersen, Terma A/S-
Examiners:
Professor Ole Winther, DTU Compute.
Professor Serge Belongie, Cornell University & Cornell Tech, USA.
Associate Professor Robert Jenssen, The Arctic University of Norway.
Chairperson at defence:
Associate Professor Anders Bjorholm Dahl, DTU Compute.
Everyone is welcome.