Connectionists: PhD Studentship on Deep neural networks. University of Bristol

[Roland Baddeley]

PhD Studentship: Deep neural networks, investigating what they compute and
why they work so well
University of Bristol - School of Experimental Psychology

Funding amount: Fees and Stipend for 3.5 years
Closes: 28th February 2017

The project: A fully funded (EPSRC) PhD studentship is available at the
University of Bristol, UK, to study deep neural networks, investigating
what they compute and why they work so well.

Deep neural networks have proved tremendously powerful methods for solving
many pattern recognition problems. They consist of a method for
representing a problem, together with techniques for finding good
solutions. In the abstract a neural network represents a problem as a
(potentially very large) number of transformations, computed by layers of
"neurons", that each operate on the output of the previous layer. The final
layer consists of a representation, knowledge of which allows the problem
to be solved. Training rules then consist of discrete approximations to
stochastic differential equations, with these noisy equations being based
on large data sets of examples of the desired behaviour. Empirically, we
know that these methods work very well and, at least when trained with very
large data sets, produce solutions that generalise better than almost all
alternative techniques. They have therefore become the basis of many of the
advances in artificial intelligence. What is !
 far less clear is why. This PhD aims to try to work out: why neural
networks work; what all the tuning parameters "mean"; if there are better
ways of parameterising these transforms other than the crude but effective
methods currently used; and if there are alternative cost functions that
result in more robust learning early on in "training".

The plan is to use methods from statistical physics and differential
geometry to characterise both what each layer/transform does, and what the
combined effect of the multiple layers is. The candidate is therefore
expected to have a strong mathematical background, and will be expected to
have taken courses in either some kind of differential geometry,
statistical physics or both. The project will require empirical
investigation of the behaviour of deep neural networks, usually using
synthetic problems where the characteristics of the problems are well
understood.  Therefore some experience in programming (any language) would
be an advantage, but given that the tools available are so good and easy to
learn, not essential.

The candidate would be based in the School of Experimental Psychology at
the University of Bristol, but we have strong connections with other
Schools within the University.

For more information contact:
Roland Baddeley (roland.baddeley at bristol.ac.uk) or
Nick Scott-Sammuel (N.E.Scott-Samuel at bristol.ac.uk)


How to apply:

Please make an online application for this project at
http://www.bris.ac.uk/pg-howtoapply. Please select PhD Experimental
Psychology on the Programme Choice page. You will be prompted to enter
details of the studentship in the Funding and Research Details sections of
the form.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.srv.cs.cmu.edu/pipermail/connectionists/attachments/20170213/c052d3ed/attachment.html>