Multi-view learning with Gaussian Process Latent Variable Models

The aim of this research project is to develop methodology for data-efficient and interpretable modeling from
multiple views. The focus will be on nonparametric Bayesian methods such as models based on Gaussian
processes. The projects goals are to develop machine learning that will be applicable to a large range of
different scenarios. We will initially focus on applications which use motion capture data. Our motivation for
this is twofold. Human motion is something which is readily interpretable by humans, making it easy to attach
"meaning" to results. This means that we can evaluate our first goal, creating interpretable models. Secondly,
we wish to be data efficient. Motion capture data is expensive to collect and requires specialist equipment.
Therefore we will have to learn from small amounts of data which requires us to use it in its most efficient
form. This also leads to the multi-view motivation, using motion capture data as an example, given that we have
several people performing the same action we will see each of these as views of the same underlying concept
directly expanding our dataset compared to if we had to learn from only one person.
Even though motivated by motion capture data multi-view learning is applicable to a large range of scenarios.
Virtually everything can be seen from more than one perspective. Such as a car driving on the street seen by
two people from different angles; or just by the two eyes of a single person. Another perspective of the car is
the sound of it, or even the smell of it. A face from the perspectives of being ten different individuals, or in
twenty different lighting conditions. Perspectives are not limited to being of physical objects; it can also be e.g.
the action of walking from the perspectives of being carried out by different individuals or groups of people. Or
a song from the perspectives of being performed by different musicians. Multi-view learning is focused on
exploiting these connections in the data to uncover latent concepts that explains each of the views in a joint
manner. This allows us to understand in what aspects the different views are similar and where they differ; as
well as conducting inference between them. A central aspect of any learning systems is to be able to interrogate
a model. What has been learnt? How likely is the model? What is the certainty of predictions? Besides
understanding being the essence of many applications, such as analysis or diagnosis, it is vital for trusting the
result of learning, the assumptions made, predictions as well as for guidance and comparison useful for further
development. Over the last decade we have seen machine learning successfully applied to a large range of new
domains. Many of its successes comes from availability of large data sets. In many ways, large amounts of data
have reduced the demands of learning systems as they can be allowed to be less abstract. However, for many
applications large datasets are not (and will most likely not be) available which means we need to use the
information available in the data as efficiently as possible. Therefore we will work on data efficient models that
use principled uncertainty propagation in order to reduce the need for large data sets.