Skip to main content

Learning and Evaluating the Geometric Structure of Representation Spaces

Time: Mon 2022-06-13 15.00

Location: F3, Lindstedtsvägen 26 & 28, Stockholm

Video link:

Language: English

Subject area: Computer Science

Doctoral student: Petra Poklukar , Beräkningsvetenskap och beräkningsteknik (CST), Robotik, perception och lärande, RPL

Opponent: Professor Søren Hauberg,

Supervisor: Danica Kragic, Robotik, perception och lärande, RPL, Centrum för autonoma system, CAS

QC 20220523


Efficient representations of observed input data have been shown to significantly accelerate the performance of subsequent learning tasks in numerous domains. To obtain such representations automatically, we need to design both i) models that identify useful patterns in the input data and encode them into structured low dimensional representations, and ii) evaluation measures that accurately assess the quality of the resulting representations. In this thesis, we present work that addresses both these requirements, where we extensively focus on requirement ii) since the evaluation of representations has been largely unexplored in the machine learning research. We begin with an overview of representation learning techniques and different structures that can be imposed on representation spaces, thus first addressing i). In this regard,we present a representation learning model that identifies useful patterns from multimodal data, and describe an approach that promotes a structure on there presentation space that is favourable for performing a robotics task. We then thoroughly study the problem of assessing the quality of learned representations and overview the pitfalls of current practices. With this, we motivate the evaluation based on analyzing geometric properties of representations and present two novel evaluation algorithms constituting the core of this thesis. Finally, we present an application of the proposed evaluation algorithms to compare large input graphs.