I've recently completed a masters thesis on EEG based mind reading, and I think I have a fairly good grasp on the state of the art in this field. I also have a copy of Kurzweil's The Singularity is Near by my bed, and I'm usually strongly optimistic about technology. But if IBM are talking about EEG based technology here, I would have to bet that they are flat out wrong on this one. I'll explain why.
Something like moving a cursor around by thinking about it, or thinking about making a call and having it happened requires a hell of a lot of bits of information to be produced by the brain computer interface. With the current state of the art we can distinguish between something like 2-6 classes of thoughts sort-of reliably, and even then it's typically about thinking of particular movements, not "call mom".
Importantly, what most people look for in the signal (the feature in machine learning terms) are changes in signal variance. And there are methods to detect these changes that are in some sense mathematically optimal (which is to say they can be still be improved a little bit, but there won't be any revolutionary new discoveries.) There may be other features to look for, but we wont be getting much better at detecting changes in signal variance.
Some methods can report results like a 94% accuracy over a binary classification problem. Such a result may seem "close to perfect", but it is averaged over several subjects, and likely varies between for example 100% and 70%. For the people with 70% accuracy, the distinguishing features of their signals are hidden for various reasons. And this is for getting one bit of information out of the device. Seems like such a device would need to work for everyone to be commercially successful.
In computer vision we have our own brains to prove that the problems can be solved. For EEG based brain computer interfaces, such proofs don't exist. There are certain things you probably can't detect from an EEG signal, meaning the distinguishing information probably isn't there at all. I'm easily willing to bet IBM money that who I would like to call can not be inferred from the electrical activity on my scalp. (Seriously IBM, let's go on longbets.org and do this.)
Thanks for the interesting detailed information about EEG resolution, which I can attest is in accordance with what I have read about neuroelectrical interaction in other contexts.
But what is to me implausible about thinking "phone Mom" and having my computer do it for me is that this scenario envisions an unusually high degree of usability that no consumer-facing software writers have ever achieved. Right now, on a BRAND NEW computer system using mostly application programs recommended by Hacker News readers (for example, I am using Chrome to Web browse), I can't count on my computer doing what I want even if I have my hands on the keyboard or a hand on my mouse. User-interface design appears to be HARD--or at least, it is rarely done right--so I am very doubtful that in five years or even twenty-five years I'll be able to use a computer that really does what I think.
Given the above comment I would be even less optimistic about a computer being able to determine the difference between thinking "I should call mom at some point" or "What a nice phone call I had with mom yesterday" and "call mom right now".
This would take the "butt-dialing" phenomenon to disturbing new levels.
> I'm easily willing to bet IBM money that who I would like to call can not be inferred from the electrical activity on my scalp. (Seriously IBM, let's go on longbets.org and do this.)
Be more precise. If you can distinguish two classes, you can convey anything through EEG, however slowly (think 'bits'). With error, use error-correcting codes and you can get relatively high accuracy.
Pedantics aside, I agree in sentiment. Even with invasive techniques we can only roughly decode movement.
Sure. Let's say a five second EEG segment, recorded from up to 200 electrodes. I bet that by 2017 we cannot accurately detect who a person would like to call out of phonebook of 50 people. Specifically, in a setting where each person is as likely to be called (equal prior probability), I bet the detection accuracy will not exceed 4% on average with a phone book of 50 people. I'm talking about the BCI understanding a "call mom" thought, not detecting it through some other means like a movement (though I don't expect that to work by then either).
My point was more about precision when making a bet, but yeah like I said I agree in sentiment. That being said, two-bit encoding is exactly what Stephen Hawking uses with his thumb, so it's not inconceivable to use it on a neural level as a last resort, however impractical. And in some cases, it is: often paralyzed individuals use their tongue to manipulate a cursor, but this can cause all sorts of problems like abnormally large tongues due to muscle growth.
And I work in a lab that does BMI work, and we couldn't do the "call mom" command in the sentiment of lars, even though we use more spatially precise recordings (multi-electrode chronically implanted arrays). So I'm with that. OTOH we can do some cool things like control a computer cursor or tv remote with motor commands like "left-up" etc. Subjects reported that after a while they would cease to "translate" thoughts from movement commands into "BMI" commands like "change channel". It stands to reason they might be able to do the phone-book thing in that case.
Of course, few people find it worthwhile to get chronically implanted electrodes placed in their motor cortex, soooo.
Why would "call mom" be more difficult than "Left left up up"? I'm guessing left, right, up, down, would map to certain EEG patterns. Why would it be more difficult to map call and mom to their patterns?
It seems to me that if you can map patterns for four directions, you should also be able to map patterns for 50 different phone book entries and several verbs.
I'm not being clear, I don't think that thinking the words "left left up up" would be detectable through EEG.
When I say detecting movement, I mean things like imagining moving a hand, a foot or a tongue. These movements use distinct areas of the brain so you can distinguish between them by looking at where on the scalp the change occurred. This is done in ways that are known to be close to perfect.
However, you probably couldn't use scalp location if you wanted to distinguish "call mom" from "call John", as they would presumably activate the same area of the brain. There are of course other things one could look at, and I obviously can't prove that it can't be done. But at the same time I have never seen any kind of positive result for an EEG classification task at this level of detail.
Well all you would really need are some easily detectable signals. So if move hand, move foot, move eyes, etc are all easily detectable then you have the basis for an interface. After that it's just a matter of building the interface around those limitations. It wouldn't be mind reading (not even close), but it seems like you should be able to get a reasonably good UI going.
I doubt it. Firstly, you will have to provide a method to filter out real movements from intended ones. A sensor on a few muscles may help, but sticking them on your skin every morning would not help towards the goal of "Reasonably good UI".
Secondly, I am not sure one can learn to almost unconsciously think about certain movements of bodily parts. Chances are this will keep requiring too much apof one's attention.
Thirdly, I think temporal resolution will be awful. Even if you can learn to think about say 3 movements simultaneously, I doubt you will get this above a byte per second of bandwidth. Written text is around a bit/character, so that would likely be way below slow speech.
Most of this is opinion/guessing, so feel free to correct things.
Rather than thinking "call mom," could you think "imagine moving your arm in that direction and pushing something"? Instead of literal mind-reading, have essentially a touchscreen without physical touch?
It definitely isn't possible now, and I wouldn't expect it in five years either. If you look at [1], you can see the areas of the motor cortex. With todays methods we can do an acceptable job at separating for example hand from foot movement. These methods look at the spatial domain, and do so in a way that is near perfect. And as you can see, there is a certain distance between the areas on the scalp, while the fingers are all in the same area.
So you couldn't distinguish individual fingers with todays technology. If it was ever to be done, I'd expect that it would be done with the same algorithms as we use today, but with much denser electrodes. If I were to bet, I'd bet that this would be physically impossible, but I'm not as confident as I am with saying we wont be able to detect who I want to call.
HA. Error correcting codes? You realize that means the person thinking in bits would have to do that themselves. A 7-bit parity would be harder to do than just dialing the damn thing. Even then all it'd be able to do is tell you that you messed up. Something that actually makes corrections on the fly (Hamming). No way. I couldn't do it and I know how Hamming works. I really doubt a consumer could do it or would want to.
A credit card number already has a check digit. If you could make people memorize a few more digits than the telephone number that they'd like to call, you have your error correction.
(Of course, this is highly domain-specific and not very convenient. But it does seem possible.)
"who I would like to call can not be inferred from the electrical activity on my scalp"
If I read the blog post correctly the claim is not passive mind reading. The claim is that the user has some training to issue the sorts of thought commands that can be reliably picked up. If I think "call mom" it doesn't do anything. But if I think "Up Up Down Down Left Right Left Right" maybe it can interpret that correctly as my preassigned shorthand for "call mom".
They're a little fuzzy, but the exact quote from the video is: "IBM scientists are researching how to link your devices, such as a computer or a smart phone. So you just need to think about calling someone, and it happens." From the context it is implied that this will happen in 5 years. In their in-depth blog post about the topic, they say: "...I could wonder what the traffic will be like on the way home and this information would pop up in front of me."
So I think they are actually describing a "call mom" scenario, and I strongly doubt that the information to detect that is at all present in an EEG signal.
Maybe if it recognized just thinking about that sequence, but I think having to deliberately focus on each direction (kind of like pressing a series of buttons in your mind, one after the other) would be too hard to accidentally do like that.
Even with very invasive techniques (chronic implants), the reading of motion signals for example is very problematic. One of the first devices to do that is the http://en.wikipedia.org/wiki/BrainGate . It requires substantial training and retraining because of brain plasticity. Still, these are major breakthroughs, and hopefully we will be able to find a reliable and usable signal, but i agree, EEG is too bulky for that purpose.
Something like moving a cursor around by thinking about it, or thinking about making a call and having it happened requires a hell of a lot of bits of information to be produced by the brain computer interface. With the current state of the art we can distinguish between something like 2-6 classes of thoughts sort-of reliably, and even then it's typically about thinking of particular movements, not "call mom".
Importantly, what most people look for in the signal (the feature in machine learning terms) are changes in signal variance. And there are methods to detect these changes that are in some sense mathematically optimal (which is to say they can be still be improved a little bit, but there won't be any revolutionary new discoveries.) There may be other features to look for, but we wont be getting much better at detecting changes in signal variance.
Some methods can report results like a 94% accuracy over a binary classification problem. Such a result may seem "close to perfect", but it is averaged over several subjects, and likely varies between for example 100% and 70%. For the people with 70% accuracy, the distinguishing features of their signals are hidden for various reasons. And this is for getting one bit of information out of the device. Seems like such a device would need to work for everyone to be commercially successful.
In computer vision we have our own brains to prove that the problems can be solved. For EEG based brain computer interfaces, such proofs don't exist. There are certain things you probably can't detect from an EEG signal, meaning the distinguishing information probably isn't there at all. I'm easily willing to bet IBM money that who I would like to call can not be inferred from the electrical activity on my scalp. (Seriously IBM, let's go on longbets.org and do this.)