RESEARCH ASSISTANT POST

[John Shawe-Taylor]

   ---------------------------------------------------------------
                   RESEARCH ASSISTANT POST
   Royal Holloway/London School of Economics, University of London
   ---------------------------------------------------------------

Jonathan Baxter has been working at Royal Holloway and London School
of Economics on an EPSRC (a UK funding council) funded project entitled 
`The Canonical Metric in Machine Learning'. Attached below is an abstract 
of the project which has a further year to run. Jonathan is resigning 
from the project to move to the Australian National University, where he 
will continue to work along similar lines. EPSRC have given us permission 
to recruit a replacement to start any time between 5th July and 5th 
January 97 and to run for a further 12 months provided that they are 
suitable for the work involved.  If you know anyone who would be 
interested, it would be very helpful if they could visit London before 
Jonathan leaves on 5th July. Jonathan, Martin and I will be attending 
COLT and so could discuss the project in detail with anyone interested at 
that time.  The rate of pay is 19600 pounds pa paid half through each 
institution. The slant taken could be towards more implementational work 
or alternatively (and perhaps preferably in view of the funding committee 
being mathematical) more theoretical. Anyone with an interest should not 
hesitate to contact us for more information.

Best wishes
John Shawe-Taylor, Martin Anthony and Jonathan Baxter
----------

Canonical Metric in Machine Learning (Abstract and progress) 

The performance of a learning algorithm is fundamentally limited by 
the features it uses. Thus discovering sets of good features is of 
major importance in machine learning research. The principle aim of this
project is to further our theoretical knowledge of the feature discovery 
process, and also to implement practical solutions for feature discovery 
in problems such as character recognition and speech recognition. 
 
The theoretical aspect of the project builds on work by the previous 
research assistant (Jonathan Baxter) showing that if a learner is 
embedded within and environment of related tasks then the learner can 
learn features that are appropriate for learning all tasks in the 
environment. This process can be viewed as "Learning to Learn". One can 
also show that the environment of learning problems induces a natural 
metric (the "canonical metric") on the input space of the learner. 
Knowledge of this metric enables the learner to perform optimal 
quantization of the input space, and hence to learn optimally within the 
environment. The main theoretical focus of the project is to further 
investigate the theory of the canonical metric and its relation to 
learning. 
 
We are currently applying these theoretical ideas to the problem of 
Japanese character recognition and so far we have achieved notable 
success. The practical part of the project will be to continue these 
investigations and to also investigate applications to speech recognition. 
 
Further information on the background material to this project may be found
in Neurocolt technical reports: 95-45 95-46 and 95-47. Also see Jonathan
Baxter's talk at this year's COLT.