In other words, ignore the PR about downloading thoughts and dreams. We can decode some stuff from fMRI activity patterns but we're soon swamped by noise, even when using machine learning techniques. The data just isn't there. It's like trying to predict the weather two weeks from now. The system is inherently noisy.

http://www.sciencemag.org/cgi/reprint/310/5756/1963.pdf?maxt...

Still, that's a big difference from predicting (or mindreading) the particular item.

And here's a good review of that and similar work:

http://polyn.com/struct/NormEtal06_TICS.pdf

Here are two empirical reports: http://www.jneurosci.org/cgi/content/full/26/18/4917 http://www.jneurosci.org/cgi/content/full/27/14/3790

And here's a review: http://www.psych.upenn.edu/stslab/assets/pdf/TS_CONB05.pdf

Happy to answer questions at grob AT mass inst tech

This, while not directly relevant, might be helpful: http://www.scholarpedia.org/article/What_and_where_pathways

Also, they are using fMRI activations, which means they don't record from neurons directly, but look at cerebral blood flow (which is linked to neuronal activation; the more the neuron does, the more oxygen it needs).

I imagine it would take a bit of training to be able to visualize images in ways that a computer can read them - I imagine children would be best at learning thought-commands quickly. Just as my grandparents could never quite learn how to touch-type, I imagine my grandchildren will kick my ass at thinking into their computers ;)