James Miskin

Who am I ?

I was a postgraduate student of the Inferential Sciences. From October 1997 to December 2000 I worked with Dr. David Mackay in the Cavendish Laboratory, Cambridge. My research is into the development of Independent Component Analysis.

In June 97 I graduated in Experimental & Theoretical Physics from the University of Cambridge. During my degree I did a Part III project with Dr. David Mackay on Independent Component Analysis. This opened up the world of the Bayesian to me and I have been hooked ever since.

I am also a keen programmer (with experience in Basic, C, Fortran 77, Fortran 90, MatLab and Visual Basic). This lead to me spending two summer vacations (1996 and 1997) working for PowerTechnology (the research centre of PowerGen) on their PROATES powerstation modelling system.

I am now working for McLaren International.

Publications

• H. Lappalainen and J. W. Miskin. Ensemble Learning. Advances in Independent Component Analysis (Ed. by Girolami, M). Springer-Verlag Scientific Publishers. To Appear, July 2000 postscript
• J. W. Miskin and D. J. C. MacKay. Ensemble Learning for Blind Image Separation and Deconvolution. Advances in Independent Component Analysis (Ed. by Girolami, M). Springer-Verlag Scientific Publishers. To Appear, July 2000 postscript
• J. W. Miskin and D. J. C. MacKay. Application of Ensemble Learning to Infra-Red Imaging. Proceedings of the Second International Workshop on Independent Component Analysis and Blind Signal Separation pp 399-404 postscript
• J. W. Miskin and D. J. C. MacKay. Ensemble Learning for Blind Source Separation ICA: Principles and Practice. Cambridge University Press. To Appear. postscript
• J. W. Miskin. Ensemble Learning for Independent Component Analysis. Thesis. postscript

Software

The Train Ensemble library demonstrates the principles of Ensemble Learning when applied to blind separation problems (as used in some of the papers above). The library is supplied as a set of MATLAB files. Refer to the README file for the list of files and refer to each example to see how to use the code.

The code works by defining a hierarchical model for the observed data given a set of parameters. For instance the ICA model uses

D=A'*s.

Each class of variable is given a prior and the algorithm proceeds to learn the optimum separable posterior distribution for all of the parameters. The modular nature of the code means that further models and priors can be included.

Currently the following priors can be applied to each parameter (including mixtures of an individual type)

• Gaussian
• Exponential
• Rectified Gaussian

The following models can be used to represent the data

• ICA
• Deconvolution of images
• Blind deconvolution of images

Further priors (binary, complex, etc) and models (non-linear ICA, etc) will be added when I have decided that they are stable.

A few related pages.

People

• My parents homepage.
• Dr. David MacKay, my PhD supervisor.
• Glyn Furlong.

Other ICA People

Interests

• B&H Jordan Mugen Honda
• Cambridge University Bowmen homepage
• Selwyn College homepage
• The Dilbert Zone

Last update: 28 November '00.
James