4 ms·

I find it frustrating when articles start out by misrepresenting the technology with "that can read people’s minds". It's understandable why they simplify thing

I find it frustrating when articles start out by misrepresenting the technology with "that can read people’s minds". It's understandable why they simplify things like that, but as someone who was involved with general EEG work and LFP work in the speech-motor cortex it feels like an unfair comparison. That's not to say that observing muscle movement isn't pretty rad, but they're different technologies reading different types of biosignals.

Natsu 9y ago
Can you provide more perspective on this & how well it can likely recognize subvocalized words? I'm wondering if someone won't find a way to do this remotely. I'd like to think that impossible, but they managed to do some very odd things like finding heart rates from pictures, so I can't really say that they couldn't somehow find a way to pick up the signal remotely and I'm curious how good the current tech is from someone who as actually used it.
- fundamental 9y ago
  Sure. Looking (quite) quickly at the paper there's a few things that stand out. - Small vocabulary sizes (20 or less words at each decision level) - A lack of discussion as to identifying utterance onset - A lack of trials of the same subject over many different days (EEG/EMG/LFP can vary considerably across subjects and within the same subject in different trials) - A lack of discussion on the cross-validation protocol (the setup of training and testing folds can provide radically different views on computational neuroscience sorts of results) All of this is fairly normal within this sort of work, however it also means that a hefty grain of salt should be taken when viewing the work. With the results that they have you have 90% accuracy when picking 1 out of 5-20 words if you setup the system and trained it the same day. That's pretty cool for small muscle movements, though not really a solution that's directly usable IMO. (also don't believe the results that are claimed by [6].) > a way to do this remotely. i.e. with a standard camera at a standard resolution/framerate? There's coarse information that you can likely see from another human yourself if you see someone mutter something, but there's too much to muddle things for it to be useful immediately. If you get to the point where you're using high resolution high framerate multi-spectral cameras you might be able to get marginally more information, but nothing too major IMO. The MIT heartrate enhancement worked well due to the periodic nature of heart rates as well as the proximity of blood vessels to the surface of the skin. Neither of those help out for observing a-periodic muscle movement under a lot of skin.
  - Natsu 9y ago
    This was incredibly informative and I just wanted to say thank you for providing a lot of perspective on this because a single upvote wasn't enough.

Natsu 9y ago

Can you provide more perspective on this & how well it can likely recognize subvocalized words? I'm wondering if someone won't find a way to do this remotely. I'd like to think that impossible, but they managed to do some very odd things like finding heart rates from pictures, so I can't really say that they couldn't somehow find a way to pick up the signal remotely and I'm curious how good the current tech is from someone who as actually used it.

fundamental 9y ago
Sure. Looking (quite) quickly at the paper there's a few things that stand out. - Small vocabulary sizes (20 or less words at each decision level) - A lack of discussion as to identifying utterance onset - A lack of trials of the same subject over many different days (EEG/EMG/LFP can vary considerably across subjects and within the same subject in different trials) - A lack of discussion on the cross-validation protocol (the setup of training and testing folds can provide radically different views on computational neuroscience sorts of results) All of this is fairly normal within this sort of work, however it also means that a hefty grain of salt should be taken when viewing the work. With the results that they have you have 90% accuracy when picking 1 out of 5-20 words if you setup the system and trained it the same day. That's pretty cool for small muscle movements, though not really a solution that's directly usable IMO. (also don't believe the results that are claimed by [6].) > a way to do this remotely. i.e. with a standard camera at a standard resolution/framerate? There's coarse information that you can likely see from another human yourself if you see someone mutter something, but there's too much to muddle things for it to be useful immediately. If you get to the point where you're using high resolution high framerate multi-spectral cameras you might be able to get marginally more information, but nothing too major IMO. The MIT heartrate enhancement worked well due to the periodic nature of heart rates as well as the proximity of blood vessels to the surface of the skin. Neither of those help out for observing a-periodic muscle movement under a lot of skin.
- Natsu 9y ago
  This was incredibly informative and I just wanted to say thank you for providing a lot of perspective on this because a single upvote wasn't enough.