Class-specific measures: CMI and CD

(This page overrides the initial work on functional brain clustering. More details about the relevant experiment can be found in Brain functional distinguishability measure.)

Here we have two class-specific measures:

  1. class-specific mutual information (CMI)
  2. class-specific distinguishability (CD)

The CMI is obtained by tweaking the traditional MI. One big disadvantage of the CMI is that the measure is not normalized and hence sometime ambiguous when compared across the scenarios. Therefore, we need to come up with some similar measure with its value normalized, CD.

We have the pre-computed CMI and CD available for Haxby 2001 data (roi = VTC, 577 voxels).

/share/Bot/Research/mvpa_data/matlab_format/gamma2p5_wholebrain/specificity_measure

Here is the comparison between CMI and CD

Next, we will plot the CMI and CD for each class on the brain. For the figures below, the color ranges from min to max in each figure only, so cannot compare across the figures. The examples of each class are shown here.

Now we will limit the value of CD from 0.05 to 1, and CMI form 0.01 to 0.1. (transparent = 0.6 on the color, = 105 on the standard orientation)

We also use the hierarchical clustering to see the simlarity among the classes. We found that all tools are grouped together. And we have special cluster for house and human face. Please see functional brain clustering page.

Next, we will use both CD and CMI the measure as a measure to cluster/segment the brain into small regions.

CD

And this is for CMI

That's pretty difficult to see the cluster, so we are working on this.

We also want to compute the similarity among classes

CD The cluster here is very meaningful.

CMI

See: clustergram in MATLAB, demo

Motivation: Mutual Information is a useful measure for classification because..., but it does not tell much about which region in the brain is maximally responsible for which class.

Objectives:

  1. See if each region in the brain is maximally responding to some certain categories?
  2. Can we use that class specific region as a kernel to help the classification?

Process:

  1. Using class-specific MI to calculate the the likelihood of the class
  2. Use GMM to model each class-specific map
  3. Use each GMM as a kernel
  4. Train/test the data using the class-specific kernel