Supervision: Ali Hariri
Project type: Semester project (bachelor)
Semester project (master)
Background: Representation learning of both structured and graph data is mostly performed in Euclidean space. Earlier studies have shed light on subspace learning methods for multi-view images. The latter can be approximated by linear subspaces, which form a curved Riemannian manifold called the Grassmannian manifold. Hence numerous studies later used Grassmmann kernel techniques for video classification and action recognition. This work explores deep Grassmannian learning methods on both euclidean and non-euclidean structures.
Project: This semester project will tackle the combination of deep learning techniques and subspace methods to be applied on both video and graph datasets. The main challenge is to learn discriminant techniques to repel samples belonging to different classes in latent space while attracting samples from the same class. To proceed, we can inspire from earlier work done on Discriminant Analysis on manifolds. Some potential datasets to be used are the NTU-RGB action recognition datasets and HMDB video datasets. Yet, the choice of datasets is flexible and can be discussed upon the start of the project. If successful, this project can be extended to numerous applications and can lead to publications.
Required skills: Previous experience with Deep Learning projects using PyTorch or Tensorflow is highly preferable. Some of knowledge of OpenCV for video pre-processing is a plus but not necessary at first, the student will be given time to acquire this skill. Willingness to read about manifold learning and optimization is a plus. We might not necessarily go into the hardcore maths of Grassmannians but it would be nice to have a high level idea at least.
Supervisor: Ali Hariri. I am a ML researcher with experience in geometric deep learning and its applications. You can reach me at firstname.lastname@example.org for any questions about the project.