MASH: Phd and post-doc positions in machine learning


The MASH project is a three-year research initiative which brings
together five institutions with expertise in statistics, machine
learning, goal planning and computer vision to investigate the
collaborative design of complex hand-designed priors for machine

MASH is funded by the Information and Communication Technologies
division of the European Commission, Cognitive Systems and Robotics
unit, under the 7th Research Framework Programme.

Research will start in January 2010 and will be carried out in
Switzerland (IDIAP), France (CNRS and INRIA), Germany (WIAS) and
Czech Republic (CVUT). Open positions are listed below.

You can already register on to get
updated by mail on the status of the project.


The goal of the MASH project is to create new tools for the
collaborative development of large families of feature extractors.
It aims at starting a new generation of learning software with
great prior model complexity.

The project is structured around a web platform which will be open
to external contributors early in 2010. It comprises collaborative
tools such as a wiki-based documentation and a forum, and an
experiment center which runs and analyzes experiments on a
continuous basis.

The applications targeted by the project are classical vision
problems, and goal-planning in a 3D video game and with a real
robotic arm.

Contributors will participate to the project by uploading the
source codes of “feature extractors” into the platform. Each one of
these extractors processes an input image to generate values
relevant to the system. This purposely broad definition spans from
classical vision processing such as edge detector or color
histogram estimation, to highly dedicated hand-designed templates
or event-based memory for the robotic applications. The system
concatenates all these extractors to create a very large feature
vector, which is used as an input signal for a machine learning

In practice, anybody can upload such a module at any time. It will
be immediately compiled and integrated in the next starting
experiment. Preliminary performance measures will be provided in a
matter of minutes, and complete results a few hours later. The
system encourages contributors to improve upon the work on other
and focus on the main weaknesses of the overall system.

The scientific issues to be tackled along the course of the project
are numerous, from standard machine learning questions such as
learning and prediction with very large feature spaces and tight
computational constraints, to original problems related to
clustering in a functional space.


– Idiap Research Institute, Switzerland (IDIAP)

– Centre National de la Recherche Scientifique, France (CNRS)

– Weierstrass Institute for Applied Analysis and Stochastics,
Germany (WIAS)

– Institut National de Recherche en Informatique et en Automatique,
France (INRIA)

– Czech Technical University in Prague, Czech Republic (CVUT)


Contact point: Dr. François Fleuret,
francois.fleuret (at),

On-line application at

The selected candidate will be a doctoral student at EPFL EDEE
doctoral school. Research will be done at the Idiap Research
Institute, under the supervision of Dr. François Fleuret.

The research to be carried out will be the study of prediction
techniques for goal-planning with very large feature space. The
candidate will investigate prediction from images, mimicking to
learn policies provided by human operators, and extensions of
classical Markovian Modeling to the specificity of the MASH

This work will mix theoretical developments in statistical learning
with the implementation of algorithms working on real-world data.

Applicants must have a strong background in mathematics and be
self-sufficient in programming. They must be familiar with several
of the following topics and interested in all of them:
probabilities, applied statistics, information theory, signal
processing, optimization, algorithmic, and C++ programming.


Contact point: Dr. Yves Grandvalet,
yves.grandvalet (at),

We have open PhD and PostDoc positions to develop clustering and
block-clustering algorithms that will summarize heuristic behaviors
across tasks. We aim at providing feedback to the heuristic
designers by detecting similar heuristics across similar tasks,
thus empowering designers to analyze coexisting strategies, and to
detect critical failures.

We will develop clustering and block-clustering methods based on
probabilistic models and factorization techniques. We will also
study the relationships between these approaches.

The candidates will hold a Master/PhD in applied mathematics or
computer science, and should have interest in both areas. They will
work under the supervision of Y. Grandvalet and G. Govaert at the
Heudiasyc lab. at University of Technology
of Compiègne


Contact point: Dr. Gilles Blanchard,
gilles.blanchard (at),

The research will be carried out at the Weierstrass Institute,
Berlin, under the supervision of Dr. G. Blanchard; the selected
candidate will be a doctoral student at the Humboldt University,

The research will concentrate on theoretical and practical
developments of prediction techniques from a large set of
heterogeneous features: aggregation, sparsification, grouping and
reduction techniques, in particular under a strong limitation
constraint of the computational burden. Automated construction of a
similarity or distance measure between features will be also

Specific Requirements: university degree (at least master/diploma)
in mathematics, computer, science or engineering. We expect from
potential candidates very good programming skills (C++) and at
least basic knowledge in mathematical statistics, theory of machine
learning and/or optimization.


Contact point: Dr. Olivier Teytaud,
olivier.teytaud (at),

The research will be carried out at the LRI, Université Paris-Sud,
under the supervision of Olivier Teytaud (INRIA research
fellow). We have open PhD and PostDoc positions.

The research will focus on theoretical and practical developments
of planing techniques from a large set of heterogenous features.

Specific Requirements: university degree (at least master) in
mathematics, computer science or engineering. We expect from
potential candidates very good programming skills (C++) and at
least basic knowledge in machine learning and/or planning.