Using independent voxel classification accuracy to determine feature order

Post date: May 25, 2012 6:50:38 AM

So far, MI between the feature and the class label have been used as a measure to determine the order of voxel fed to a classifier. However, high MI does not always imply discriminative feature. Here I use classification accuracy of one feature alone to determine the order.

We also experiment on the data and found that the order determined by MI and accuracy are related but NOT consistent. So, I'm curious to know which would work better.

A: MI-descend outperforms svm-accuracy-descend

I made MATLAB code available:

demo_calculateSVMScore.m

svmAccuracyScore.m

It's very interesting that using only single voxel to classify the data set would give very low accuracy. The best accuracy for single voxel is 30% accuracy. Recall that when we allow multiple features, we can reach 95% accuracy.

After the experiment, the answers become clearer:

1. MI-descend outperforms accuracy-descend because the poor accuracy obtained from a voxel does not mean imply the voxel is not informative.
2. Moreover, when the voxels are used jointly with others, they, as a group, perform much better as the information in the brain is distributed, consistent with the results in Haxby et al.