alexgmcm / dmeproject Goto Github PK

MATLAB Licenses

I tried using the MATLAB SVM thing just to test how the libraries work - and ran into a problem that either we don't have the bioinformatics toolbox (required for the SVM function apparently - itself in the stats toolbox) or there are insufficient licenses. This is a pain but on my laptop I have the full MATLAB w/ all toolboxes thanks to a certain Swedish website so while irritating it isn't a major problem.

Wiki

We might also want to make a wiki on github as a way to keep notes as we work on it.

First Steps

I suggest today we try to outline some of the methods we might try.

Amos suggested using a Hidden Markov Model as an explicit time series model. There is a decent paper on using them here for the asynchronous case we have - also it describes Input-Output HMM's which might be fun as they are an extensions of HMMs which in the paper already uploaded by Lotte et. al. IOHMM's outperformed HMM's on a similar protocol to what we have. Paper here: http://bengio.abracadoudou.com/cv/publications/pdf/rr03-49.pdf

Multilayer perceptrons were also very successful and there are numerous references in Lotte et al. to papers describing the implementation for EEG BCI. Lotte also suggests Support Vector Machines and Linear Discriminant Analysis. Also methods based on the Mahalanobis distance had some success.

At the moment I've kind of been focusing on PMR, and I have a research proposal to write which hopefully won't take too long. But then I can focus on DME (both this and the presentation). We have until next friday when we have to submit the interim plan thing so by then we should have decided the methods.

I'd recommend we choose one each or something and type up a chapter about how it goes, and then stick all the chapters together and add an introduction and conclusion chapter to form the final report?

Our precomputed features are the power spectral densities so if you can stick to classifiers which used those in the literature (it's a very common feature set to use) and try and use libraries wherever possible then obviously it will be a much easier task :P

See you guys in a bit anyway. I hope you are feeling better Tim!