simple classification on fMRI data

Logistic regression

In this page we will show how to use simple logistic regression to classify the fMRI data from Haxby et al. This requires two toolboxes:

• Logistic regression toolbox adapted from Prof. Ng's coursera
• Miscellaneous toolbox containing some necessary functions

You may also use this m-file as a kick-starter. The data example is preprocessed (i.e., detrended, z-scored, ) by our approach described in preprocessing toolbox. The data in mat-format is available here.

Logistic regression is naturally a binary classification algorithm, but somehow can be cast as multi-class classification algorithm by adopting one-vs-all (one-vs-rest) strategy. Another possible strategy is pair-wise classification. Personally I prefer one-vs-all over pairwise because the latter one need to classify K(K-1)/2 (K choose 2) times as opposed to K times in the former approach which is implemented in our code.

The accuracy is 77.5% when using 50% of the data instances (40/80) to train and test on the rest (40/80). The accuracy achieves 87.5% when using 72/80 to train the model and test on the rest 8/80.