Friday, March 27, 2009

EEG consumer BCIs spanked by the Schwartzinator

Forbes has a fun little story on the state of consumer BCIs that hits on the major consumer players, but also throws real BCIs researchers John Wolpaw and Andy Schwartz into the fray. Each echos the officially sanctioned position of their respective scalp electrode versus implanted microwire/fancier probe electrode camp.

I prefer to call this the Head-Computer Interface versus Brain-Computer Interface debate. I try to be even keeled on the subject. EEG is obviously a worthwhile endeavor because there's none of that messy surgery involved, and thus appeals to a very large audience (at least in the near term). Hey, if you get a signal that can be controlled by the user reliably and with an acceptable level of precision, you have something useful. BUT, EEG is a degraded, garbled, sloppy signal prone to nearly limitless interference sources. Anything that can bork an implanted BCI can bork an EEG-based system, but EEG throws the doors open to the world because you are essentially wearing an antenna (or several).

One major problem with EEG is the fallacy that somehow destroyed information can be recovered by some form of fancy filtering. This is simply not so. Think of the electrode as a point of convergence for all electromagnetic signals of a measurable intensity. Even after narrowing frequency bands and implementing funky probabilistic decoders, any squiggle can be the convergence of several squiggles of indeterminable sources. In other words, at a specific time, a signal of amplitude +5 can be two signals of +4 and +1, -2 and +7, or 5 signals of +1-4-2+3+7. And don't even start with harmonics.

The second problem is population size. I know we like to think that motor cortex responds to only movement. It makes life easier. What life? Life in La-La-Land. MI responds to visual stimuli, movement preparation, auditory stimuli, imagined movements, movement related words (heard, internally rehearsed or spoken), reward, attention, cutaneous and proprioceptive feedback, and a bunch of other factors I'm not even mentioning. Until the impact of these influences is understood and quantified, there will not be any BCI that translates the neural activity for "move my arm to point x,y,z". EEG will never have the fidelity to isolate the differences feedback has at the single neuron level, so the nature of the recorded signal will never allow the 'direct' mapping of neural activity to output. That is, the activity that once gave rise to movement can never be harnessed with an acceptable degree of control to recreate the movement. Yes, I am using the word never. Never. There, I said it again.

Like I said, this isn't an attack on EEG, just reality. Different uses for different technologies. I can drive a car on a road. That doesn't mean I can drive a road, or that a road has no use.


Brendan said...

Mr. King,

I posted the comment below on:

I also wish to thank you for drawing attention to BCI research. While I disagree with your points and, moreover, your unnecessarily snide attitude that can only draw a proportionate response, I nonetheless support your work in presenting BCI research to the public. Good luck to you, I will be happy to buy you a beer at SfN in October or whenever we meet.

Also, it is Jon Wolpaw, not John.

My posted reply is as follows:

Thanks, Gerv and Michael, for your excellent responses. Schalk's comments are discussed in more detail in his outstanding 2008 article, Toward Brain-Computer Symbiosis. Indeed, the silly assumption that 3D movement can only stem from direct identification of x,y,z coordinates is addressed there, and numerous other articles.

This sort of passive-aggressive condescension from invasive BCI researchers is nothing new. See Wolpaw and McFarland 2004 for a response, or better yet, their more recent demonstrations of noninvasive 3D BCI control. Some invasive researchers said this was impossible too. See the famous Carmena et al 2003 paper, and Nicolelis' entertaining misstatements to popular media around then. But perhaps the best response is the 60 Minutes show that featured both invasive and noninvasive groups (including Wolpaw and Schwartz) and aired on 2 Nov, 2008. We see one user who attained success with an invasive approach, and another who used a noninvasive approach. Two very different approaches, from different groups, that both attained the same goal. Ahhhhh. Let patients be informed and choose. The show neglects the possibility of invasive and noninvasive hybrids, but that is OK, they do not exist yet.

What is most unfortunate about Mr. King's patronizing condescension is that such an attitude really stems from a minority of invasive BCI researchers. I am on friendly terms with Kai Miller, Eric Leuthardt, Dawn Taylor, Jane Huggins, Leigh Hochberg, John Donoghue, Phil Kennedy, Joe Wright, Gerv Schalk, etc. I respect them and their work, and the feeling is mutual. I hope they get more funding. Blog broadsides such as his do not help.

I visited the HQs of Emotiv and Neurosky in April 2008. I used both systems. Indeed, they are very far from 3D control. This misses the point. The goal of those systems is to make something that can be fun, and moreover supplement conventional controls rather than replace them. True, their systems (very probably) rely heavily on non-EEG info. This also misses the point. If the goal is to make a fun system for healthy users, who cares what inputs it uses?

I do agree with a bigger concern: these companies have missed deadlines and so far not lived up to the hype. But, I am reasonably confident that gaming can be enhanced by head mounted sensors. For more on this, see my IEEE article last year (Allison and Graimann, 2008, part of Nijholt et al., 2008) or other work by my colleague Prof. Nijholt.

Mr. King wrote:
I know we like to think that motor cortex responds to only movement. It makes life easier. What life? Life in La-La-Land.

"We" who? To which paper (tiger) do you refer? Can you provide a paper by an established BCI expert that says that the motor cortex responds only to movement? The fallacy of fancy filtering that Mr. King mentions is a bit more justifiable, but I think most noninvasive BCI researchers, including me, will happily agree that invasive BCIs can provide a more detailed picture of single neural activity. No question. If you happen to be interested in that neuron, or its neighbors; electrodes in M1 tell you little about visual attention, and there are obvious limits on how many electrodes the brain can handle, the number of surgeries, risk of infection, etc.

Invasive BCI research is important, exciting, and promising. If I become locked in, I would consider an invasive system of the future, ideally hybridized with a noninvasive system to provide less detailed information from areas not covered by invasive electrodes. For the overwhelming majority of the population, noninvasive BCIs are the only options.

"Your brain or mine?"
--Jon Wolpaw, opening address of the 2003 BCI Symposium at the Society for Neuroscience conference, referring to invasive vs. noninvasive methods

--Dr. Brendan Allison
Laboratory of BCIs
TU Graz

kai joshua said...

Dear Mr. King-

A few quick points:
1) The activity of a single neuron can't reconstruct movement parameters on a single trial basis, either. Because spike output is stochastic, you either have to average across neurons (using an MUA) in some way or integrate over time (as is frequently done).
2) The brain appears to represent stimuli and motor planning with populations of neurons in which each neuron's activity represents a probability distribution across stimuli/planning space (think Hubel and Wiesel). It stands to reason that "funky probabilistic decoders", as you call them, would, in fact, be a reasonable way to approach understanding neuronal activity, whether at the single neuron, neural population, or even EEG scale.
3) Your comment that: "there will not be any BCI that translates the neural activity for 'move my arm to point x,y,z'." is incorrect. The newest Wolpaw 3D videos show exactly that, using EEG. While these may only have been shown at conferences (so you may not have seen them), they are extremely convincing.


Brandon King said...

I'm not used to meeting people with the same name, so it gives me great pleasure to debate another Br(x)nd(y)n. Perhaps you are my arch-nemesis or doppleganger... but, by your profile pic, I see you wear white, and I always wear black, which means I would have to be the bad guy, and we know how that always turns out...

Thanks for pointing out the Jon misspelling (there were some other typos in there that I'll fix in a sec). I'm sorry if the post sounded condescending because that wasn't the intention. The audience for the blog is 'general' and the tone usually whimsical, so I can see where that could get misinterpreted or taken too far (by me when posting). Chicago is my stomping ground (home town), so the beer is on me!

I haven't played much with LinkedIn (used it just for my tech contacts and have just started adding academic ones), but that url goes to a 'membership required' group, so I can't see the full thread. But, let me address some of those comments quickly.

First, I'm definitely not saying that you need x,y,z endpoint control as the only means of 3D input. Totally agree there.

Second, the 3D non-invasive control, and for that matter most of the cursor control demos, use a number of interface tricks in the actual metrics measuring performance - targets the size of the screen edge, dwell click, selection only in target areas, and recentering the cursor to name a few - so direct comparisons of performance to invasive BCIs is difficult. Check out SP Kim et al in J Neural Eng (2008) (PMID: 19015583) for the direction the people I work with are heading. In general, I like to shy away from the media demonstrations and non-peer reviewed or abbreviated experiments that come out of some labs (Nicolelis seems to do this more often than some others, but everyone's guilty and there are some good papers coming out of his lab besides the monkey control robot earlobe on International Space Station via Web 2.0 interface).

One of the thoughts that drove me to post was the idea that with all the consumer devices hitting the market, from kids' toys to game controllers, this is going to be a big "put up or shut up" trial for EEG systems. I honestly don't think the products being shown off will take off because I think the research is really just finding its 'rhythm'. Not all media attention is good. When I worked with Phil Kennedy, he consulted on an episode of Chicago Hope where they used an implanted BCI to communicate with a locked-in patient. We were a little unsure what to think when we watched it on network TV and found out the first thing the patient says is along the lines of, "Please let me die." Like it or not, from now on EEG BCIs are going to be intertwined with the success or failure of a Mattel toy. That worries me. Look at the OCZ NIA feedback on NewEgg. Be sure to chop a star off that rating, too, because the marketing dollars behind it, the near stalking customer service responses, and the constant special savings, along with the wireless version in development mean OCZ is most likely padding the rating. (In fairness, OCZ does make some great stuff, and their support is very responsive, though the build quality of this product isn't up to their usual standards.)

Anyhoo, as for the 'We', I meant the more global 'We'. It wasn't aimed so much at researchers and laypeople reading the post, and while I hope my readers are beyond the 'we only use 10% of our brain' types, very very very few textbooks or discussions of neuroscience by the media (even the scientific media) make any mention that non-motor processing occurs in 'Motor Cortex'. The assumption many otherwise knowledgeable people is that Motor Cortex is called _Motor_ Cortex because it only does motor things, otherwise it would be 'Motor-Related' cortex, or 'Motor-but-more-than-that-but-it-gets-complicated' cortex or 'Motor*' cortex. So, there ya go (pilgrim).

Seriously though, I've mentioned in numerous past posts that both invasive and non-invasive BCIs are useful, intended for different users, and worthwhile avenues of research. I have also echoed your sentiments on the importance of extracting volitionally controllable signals using any means available for people that might one day rely on these systems as their sole conduit for communication.

I don't think that showcasing BCI toys that use $20 worth of parts is a good way to promote the technology at this stage. I would hope that manufacturers would at least call thee things 'biosensor devices' or something specifically NOT BCI, but I guess brain terms are sexier. If noninvasive BCIs were in the homes of even 5% of the medically-relevant user population I would think differently.

I will stand by my statement, though, that there will never be a non-invasive BCI that will allow a used to reach out, grab a bottle, and pour a glass of water using the same neural activity that once drove the same behavior. Take it as a challenge to prove me wrong. I would love to be!

I surprise I got so many panties in a bunch over this post. Actually, after rereading it just now, it _is_ damn accurate, and with the exception of the cheeky 'never' sentence, a pretty level headed assessment.

Brandon King said...


Thanks for the feedback. I'm sure some of this was a little unclear, so hopefully this answers some of your questions.

1) Yup, agree. Except I never said anything about doing this on a single trial basis. You build a decoder using a calibration task that assesses firing properties and map those back to intended output in "real time" (okay, near real time).

2) Yes, and my lab uses various methods to infer the intended output (my SfN poster two years ago used MUA and a classifier to identify intended targets, so right there with ya). BUT, we are no reconstructing the source signal. The point was that information is destroyed via interference, and the best you can hope for is to use generate the missing control signal functionally based on such a scheme. It would be as if an invasive system simply didn't record (apparently) randomly chosen spikes. A fair response would be that the stochastic nature of spiking means filtering is necessary for invasive BCIs, but my response would have been that in one case you have the signal, and in the other it has been destroyed and you have to recreate the signal accurately and then generate the output using (at least in part) the generated signal.

3) I said not until the influence of feedback on the recorded activity is accounted for. As far as EEG goes, no, there has never been anything close to the reconstruction of an accurate arm movement where the reported experience is, "I just thought about moving my arm like this and the computer moved my arm like this." I saw some demos of the 3D control you are talking about (I think) and rather than fan the flames, will abstain from responding (it wasn't bad, but I don't want to get into a debate on how good it was or the problems, etc). See my comments in the reply to Brendan on the interface tricks and metrics.

Black said...

As always, I love your posts.

Also, I love how saying "this isn't an attack on X" is the surest way to get very serious people to defend X. Call it reverse psychology for professionals. Good job!

@Other people

RELAX. The original post was hardly a "patronizing condescension" and most certainly was not a "blog broadside". You do not need to mentioned all of the researchers you personally know and respect from both camps. No one cares. These interwebs are a nasty place, but this post was soft as a kitten.

And besides, the points made about EEG are valid. It has problems. So does BCI, but the post wasn't about that.

Actually, that's not a bad idea. Brandon, let's see a post about some of the fundamental problems with the implants. A little self-reflection if you will.

Brandon King said...

Oh, thanks Black. I needed more work. Actually, not a bad idea. Now how to I extend "The biggest problem with implanted BCIs is that they might be TOO good" into a whole post... Hmm... j/k

Post a Comment