[amiKY]

Thanks for sharing. It will be interesting to read the paper, usually when you see exactly what they did the sheen comes off considerably. From what I understand, it's basically taking advantage of the doppler effect in the microvasculature to estimate brain activity? I guess the concept is nothing new (ultra sound has long taken advantage of doppler shift), it's more that they've found a way to get some images apparently with a spatial resolution that is sufficient enough to be used for "mind reading".

I did some similar experiments in my old job, but we used fMRI in the human visual cortex instead of ultrasound. We found that even when we imaged at surprisingly low resolutions (where the microvasculature couldn't be imaged) there was a still a lot of data which when we threw into a machine learning model could be used to predictably estimate the visual stimuli the person had seen during a basic task. In case you aren't aware, fMRI is actually measuring the oxygenation of the blood (not actually brain activity, we simply infer that areas with high oxygenation are highly active - which is an argument for another day!), and we theorised that the reason we could still decode information at low resolution was because of large gross veins distributed across the brain. If you imagine that when shown two stimuli A and B, the brain has two distinct patterns of activity A and B, the blood flow in those two cases will be slightly different and this will be reflected even in the gross veins, and we believed our machine learning model was able to see those difference in the gross veins and using those to "mind read".

I would definitely want to look at the study more carefully, to see what the investigators have done to remove the possibility that there are gross activity pattern changes that their learning model (I am sure they're using one) might actually be using to "mind read". It will also be interesting to see how the model is built and what they are doing to reliably clean up the data.