Connectionists: Ph.D. Project at Pattern Recognition Laboratory, Delft University of Technology

[Marco Loog - EWI]

Ph.D. Project in Dissimilarity-based Multiple Instance Learning

Pattern Recognition Laboratory
Delft University of Technology
Delft, The Netherlands

  prlab.tudelft.nl

The Pattern Recognition Laboratory invites applications for a Ph.D.
position in pattern recognition at Delft University of Technology.


Research Project

Multiple-instance learning extends classical supervised learning so as
to handle objects that are described by a set of instances, i.e.,
feature vectors, as opposed to a single feature vector only. Various
classification routines have been devised to learn from a collection of
such sets and their corresponding labels and to generalize to new and
unseen examples. Dissimilarity-based approaches -- put forward by
Pekalska and Duin [www.worldscibooks.com/compsci/5965.html] and further
investigated in for instance the European SIMBAD [simbad-fp7.eu] project
-- provide new opportunities to tackle multiple instance and related
problems. Not necessarily relying on individual feature vectors, a
dissimilarity approach opens up the novel possibility to compare sets in
a direct, albeit potentially non-Euclidean, way. The principal focus of
the research project is on developing tools and theories to carry out
multiple instance learning via such dissimilarity representations.
Additional investigation can, for example, be towards the potential
benefits of non-Euclidean, or even non-metric, representations over
[implicit] Euclidean representations like the well-known kernel
matrices. This project is fundamental in nature and the aim is to gain
knowledge and understanding applicable to a broad range of general
multiple instance problems. There is no main application involved, but
the project does link to various other research directions at the PRLab,
some of which are more applied projects we are participating in.


Qualifications

The successful candidate should have an M.Sc. in physics, mathematics,
statistics, computer science, electrical engineering, or a related
relevant discipline. A solid mathematical background, considerable
experience with pattern recognition or machine learning techniques, and
good programming skills in Matlab are definitely an advantage. Possibly
more important are the skills and drive to tackle basic, fundamental,
and/or conceptual problems. Creativity in finding solutions is thus
essential, along with good communication skills.


Applications Letters

Applications must be submitted electronically to m.loog at tudelft.nl and
be received no later than June 18, 2010. Your application should contain
the following information.

- A letter of motivation [one page maximum]
- A curriculum vitae, including three references and, possibly, a list
of publications
- Documentation of completed degrees and graduate courses, including the
marks obtained
- Detailed information on your M.Sc. project and a copy of your M.Sc.
thesis or a draft of it.

It is optional to send along a sketch, no longer than one page, of a
proposed, more detailed research direction that fits within the research
project described above.


Contact Person

Marco Loog

  m.loog at tudelft.nl

Pattern Recognition Laboratory
Delft University of Technology
Delft, the Netherlands