Both methods require an extensive training set. The only difference (that I can see) between the two is that the Japanese model assumes the representational structure of the underlying cortical areas being recorded (i.e., the size of the receptive fields), whereas the Berkeley group trains a simple linear classifier. Expect a few more of these classifier methods to come out in the next couple of years.

That's true, but advances in Optical Imaging might overcome the current imaging limitations. But sci-fi aside, at least getting the decoding algorithms perfected will answer a ton of basic science questions about network dynamics in primary sensory and motor areas.

A Berkeley group has already reported this in Nature using similar methods: http://www.nature.com/nature/journal/v452/n7185/abs/nature06713.html)