Ok... I have a real general question for anyone, what exactly are they trying to model? This was never clear in my mind (I think due to a lack of understanding of a fair amount of the terminology).

Also, I don't think the paper makes exactly clear what a "natural" image is and why only natural images are used?

Edited by author 11-08-2001 12:18 AM
I thought this paper did a very poor job of explaining the terminology used in the paper. It is not very well organized. They should have organized the paper such that it explicity and clearly described ICA and the old approach, separate from the introduction. Later parts of the paper are not well organized either.

I would be curious to know how the old approach is similar to PCA, if at all. Likewise, they should have offered a more complete analysis of comparison to other methods, and with more experimentation.

I agree with the comments by Joe that the paper was not well organized and did not explain the terminology that was used in it.

I think what they have tried to show is that using a two-layer model and enforcing the sparseness constraint on the output of the second layer a topographic organization naturally emerges in the first layer of the model.

Maybe the talk will help to better understand the point of the paper.