We are interested in widening the scope of Deep Learning, which has
proven to be a powerful tool for many problems in artifical
intelligence, in particular, the subfields computer vision, natural
language processing, and audio analysis. Geometric Deep Learning, that
is, going beyond Euclidean data, has been identified as a major
challenge. The proposed project is based on two approaches for
Geometric Deep Learning that are able to learn in a more effective way
using less annotated data.

The first one is concerned with the internal network structure and the
aim is to inject geometric properties such as invariances and priors
into the internal representation by construction. The second one uses
the fact that often unlabeled data is abundant and correlated, for
instance, through time series measurements and consequently the
network output should be correlated as well. This can be turned into a
supervisory signal for training a network in a weakly supervised


Linköping University



Educational level:

Master Degree

How to apply:

Please mention NLP People as a source when applying


Tagged as: , , , ,