Keji Li

Grandmother cell not very disprovable

One objection to localized coding (grandmother cell) is that you can't find it anyway. Well, not so, because a localized coding system will almost always manifest as a sparse coding system (or even highly distributed), given a hierarchical model.

In this example, if you are looking for Jennifer Anniston cell, but your electrodes are in stage -1 or stage 1, you will always get multiple clusters responding to your Anniston picture.

In this example, if you are looking for Jennifer Anniston cell, but your electrodes are in stage -1 or stage 1, you will always get multiple clusters responding to your Anniston picture.

"Blond" is probably a poor example here because 1) this classifier is probably still far away from the final stage of face recognition, 2) you can't mistake it to be Anniston detector, since you almost certainly have multiple blond people in your stimulus set. However, we already see the "Friends" detector working in the original paper. Furthermore, if the area they are in is actually in the same dimension as 'famous sitcom' then you can indeed find clusters that respond to "Friends".

Now, for the lesion argument, if you can't find it anyway, why would you expect to lost it?