There's been a bit of buzz about a paper that just came out in Neuron. The authors use fMRI signals from visual cortex to decode 10x10 images that subjects view while in the scanner. Since this (neural coding) is more or less what I do, I figured I'd toss in my two cents.
The idea of mad scientists "reading your mind" is probably a little scary to some people, but there's actually a lot of good that could come of it. For now, I wouldn't worry too much about people reading your thoughts - there are big limitations to taking this further.
Pharyngula points out three potential limitations, but I don't think they really get at the heart of the issue. The main limitation is resolution. fMRI can only record 2mm voxels (a 3D pixel) of activity, and the response in each of these voxels is recorded every 2 seconds or so. This is nowhere near fine enough or fast enough to do serious decoding. The fact that the part of the brain the authors record from (primary visual cortex) is laid out like the visual scene (retinotopy) shouldn't matter for decoding, but in this case, because the resolution is so low, it helps (averaging neighboring voxels can improve the signal).
The amount of computational power the authors use to decode the images isn't really a limitation at all. The math is pretty simple; I doubt if it would impress many statisticians or computational neuroscientists. More importantly the math is clean, and they cross-validate their data. They fit the model parameters on one set of data then evaluate how well it does on a different set of data. In models with many parameters it's pretty easy to get good fits. In an extreme case a model could have more parameters than data points! But because the authors evaluate the model on test data, the decoding accuracy does mean what you think it means.
The paper is a great proof of concept, though. fMRI studies have been looking at correlations between the external world and behavior for decades, but this is the first one I've seen that explicitly tests how good these signals are for reconstructing stimuli. Using the word "neuron" as an example seems a little cheesy, though. I wonder how good the decoding was for the "science" and "nature" stimuli :).
Y MIYAWAKI, H UCHIDA, O YAMASHITA, M SATO, Y MORITO, H TANABE, N SADATO, Y KAMITANI (2008). Visual Image Reconstruction from Human Brain Activity using a Combination of Multiscale Local Image Decoders Neuron, 60 (5), 915-929 DOI: 10.1016/j.neuron.2008.11.004
Posts
