Comparison k-nn and k-rn on VTC and whole brain

Objective: We hypothesize that voxels in close affinity would give better classification accuracy than voxels with random neighbors.

In this experiment, we compare the classification accuracy of both k-nn and k-random neighbor (k-rn) with the same set of seed voxels.

Results: The results ARE OPPOSITE to our hypothesis. That is, seeds with k random neighbors give better accuracy than with k nearest neighbors. This might give the insight that:

1) Information is NOT distributed locally in the brain!!!

2) Perhaps, this says the specific-category regions exist in the brain

Experiment design:

1. We first come up with a set of seed voxels S = {s_i} picked randomly in an ROI (e.g., VTC or whole brain)
2. For each seed s_i, we create 2 things knn_i and krn_i:
   1. knn_i = s_i union with (k-1) nearest neighbors of s_i
   2. krn_i = s_i union with (k-1) random neighbors of s_i
3. 10-fold classifying knn_i and krn_i and report the accuracy for each i
4. Finally, we compare the results between knn and krn in both Ventral temporal cortex (VTC) and whole brain (WB)

List of experiments:

• Experimen1: Comparison of knn and krn in VTC
• Experimen2: Comparison of knn and krn in WB
• Experimen3: Comparison of knn and krn in both VTC and WB:
  • accuracy histogram: Compute the histogram of accuracy vs portion of voxels
  • balance curve: Compare the accuracy of knn and krn point by point
  • accuracy vs k: Plot the trend of accuracy when k is varied, comparing knn and krn

Experimen1: Comparison of knn and krn in VTC

Experimen2: Comparison of knn and krn in whole brain (WB)

Experiment3: Comparison of knn and krn in both VTC and WB

Now, let's compare the accuracy from using k-nn and k-rn on both VTC and whole brain

What I observed: