So 10ish words per minute? Not blazing fast but for leading research pretty darn good! 40 WPM it probably beats the average human and most cell phone applications.
I guess the typer's experience could be a factor. I would expect someone to type around 10WPM - or even less - the first time they were placed in front of a keyboard. Getting to 40WPM usually takes years of practice. No evidence one way or another if this method could achieve the same with just practice and no changes to the technology.
> Getting to 40WPM usually takes years of practice.
30-40 WPM was the target for high school computer typing courses in the 90s (my experience, may have been a similar target earlier and I don't know the target for courses targeting typewriters), and this was a time when computers weren't ubiquitous like today. It's very feasible to hit that target in a few months especially if you learn proper hand placement (hunt & peck, two finger typists, will have a longer ramp up time but can often still hit those speed targets).
The NN in the article is based on handwriting, so comparing WPM to handwriting is a better apples-to-apples comparison than typing:
> Studies compiled by Amundson (1995) show that copying rates using handwriting at the 1st grade level are about 5 words per minute (WPM) on average, but by the end of elementary school at the 5th and 6th grade level are about 10 to 12 WPM.
From my friends who studied neuroscience, the hardest part is getting any signal from the crazy amounts of noise.
Maybe with the crazy amount of ML research poured into that space we can get one of those nice cross-disciplinary breakthroughs that come from similar tech (bluetooth/wireless sigproc, audio, etc.)
They already are, they bought CTRL-Labs which made a prototype wristband for reading motor neuron activation. They have a demo on YouTube showing it for keyboarding as well as for surprisingly accurate 3D hand pose estimation.
Purely medical implications of such tech are more profound than I can imagine. And people have been imagining it for decades.
Can you even imagine StarCraft or Quake matches, or better yey, some completely new games developed for brain computer interface? Man, it's VERY exciting.
I think the inflection point will be a game or demo that gives the experience of moving an object around with your mind. That, and then being able to summon shapes or objects just by thinking of them. I wonder if Valve has accomplished this inhouse yet. I think a VR headset with a BCI will be what kills the monitor.
As exciting as this is now, I'm very eager to see how brain-machine interfaces mature with findings how to best meet the machine half way.
Decoding hand writing accurately is an approximate thing even when it's spoon fed to the machine via a digitizer, to say nothing of interpreting the movements from neural activity. What could be accomplished if we abandoned the familiarity of writing and existing glyphs and established a novel set of learnable 'brain semaphores' (e.g. visualize this symbol, play this sound on your mental speakers) that were easy enough for the user to practice and master with feedback, and blazing unambiguous markers to the probes and recognition algorithm? It's not hard to imagine that with deliberate practice and immediate feedback, the brain can develop connections for the sole purpose of the interface that no longer even piggyback on imagining motor/sensory events. Sending a keystroke would be no different than lifting a finger.
I'm sure this isn't a novel idea, but my armchair observer prediction is that the field will start to move in the direction of purposefully designed mind-sign schemes and away from mimicry of physical movement as the technology becomes ubiquitous and less invasive.
That's an interesting idea, but what makes you confident that any concept/image/sound you hold in your mind can be less squishy and vague than traditional sounds or characters?
You've put to words what I've been thinking about for many years. I'm one of those people who find languages difficult and often find it difficult to convert my thoughts into words and sentences. From my perspective, 75 percent of mental effort goes into sentence formation and this process often derails my thought process.
When I discovered NLP models like Word2Vec and Thought-Vectors, which assign vectors to words or even whole sentences/concepts, intuitively, it felt like that was exactly what was happening in my mind. It might be an illusion but I do not think in sentences or words, I only think in concepts and images that appear in my mind's eye in an instant. To form a larger idea, I build a chain of concepts. I am sure that eventually probes can pick up a clear distinct vector that uniquely summarizes the entire concept of what I was thinking or visualizing.
And as you suggested we could start off with a smaller subset of semaphores or concepts or visualizations that at first are the easiest signals to pick up via probes. After practice, these signals only get crisper.
It seems like a direct brain-to-speech approach would make more sense for this application (communication from paralyzed people.) This is sort of like trying to write with a theremin, eh?
I'm still on the far side with this, I know, but whenever these BCI stories go by I am bemused. You connect a brain to a computer and then "train a neural network" on the computer side, that makes no sense. The sophisticated NN is on the brain-side, yes? With an integrated subjective UI already, yes? You do not need fancy implants or hardware to "talk" to your machine, galvanic response and (now) micro-IMUs (inertial measurement units, one on each finger) are plenty of bits-per-second for a subjective Think-to-Type experience. Use hypnosis, it's easy, to program your brain to output a byte by changing the angle of your fingertips and then twitching the thumb to "clock" the data. Or just standard stenotype, it would likely be easier, eh?
You don't need fancy hardware or NN software (and especially surgery to stick electrodes in your brain) to "talk" to your machine with your mind, you already have all the technology you need "built in" to your nervous system. It's not even difficult to use. Literally the first thing I learned to do when I was studying (self-)hypnosis was setting up a binary "yes/no" signal (a finger twitch) from my unconscious mind to facilitate communication.
Anyhow, not to go on and on about it, the bottom line is when you connect a computer to a brain please remember that the brain is the fancier, more powerful "device", okay? You will waste less time. (I am not sure why I am not able or willing to try to make something out of this. I can talk to my computer well enough that I don't bother with BCI. ANd somehow the plight of these poor folks doesn't move me enough to do something. Am I a moral cretin? I'm seriously you guys. Should I try (harder) to get something going with simple hypnotic BCI systems? HN?)
Maybe the connections are in reality electronic stimulants to release dopamine, serotonine, endorphin & GABA et. al. in the mix of your cocktail without using pharmaceuticals. Neural link is a big Pharma killer dressed as a pig. Use a NN in the feedback loop trained on your autocomplete output.
I totally agree. There is more than enough "unmapped" area on the motor & sensory homuncului[1] (creepy image warning) to be repurposed as an I/O channel by retraining the brain. In general, brains are plastic enough to remap senses like this[2] given a handful of months' work/practice. Adults are capable of learning ASL, for example, which uses completely different "output channels" from spoken language. Even when learning to drive, the car can start to feel like an extension of one's self (and I'm not a car person at all). Famously, people who become blind begin to develop better acuity in their other senses because of all the "freed up space" in the visual cortex (which I'm sure is an oversimplification, I'm not a neuroscientist).
We have so much raw data coming into our brains--via sight and sound especially, but surprising amounts via other senses as well--that I see little reason to add "additional senses" or outputs when we have rather high bandwidth already. Heck, my thoughts run ahead of my speech or typing fairly often, and both of those are pretty low-bandwidth channels.
One-hand chorded keyboards[3] are perennially reinvented by someone with an Arduino and a few buttons (not to belittle that work!) but you can't just pick up a chorded keyboard and type as fast as a regular one, so they are generally treated as a novelty. But if one spent as much time learning to use a chorded keyboard as they did a regular one, I'd bet they'd be just as fast. Perhaps the reason fluent computer users show little desire for better human-computer interfaces is that they are afflicted by a generalized form of baby duck syndrome[4]. To mix metaphors in a really confusing way, interacting with a computer with a keyboard, mouse, and monitor feels like using GNU nano: shallow learning curve, but it tapers off quickly. I wish I knew the Vim of HCI: I'd be willing to practice for years to adapt to it, to compute more efficiently for the rest of my life.
> I'd be willing to practice for years to adapt to it
Ah, but learn a little self-hypnosis and it won't take years.
I have a new tiny business selling custom computer peripherals: things like levers, crank-wheels, knobs, sliders, etc... for input, and odds and bods for output like sand-tables with a stylus on an arm, or lasers and mirrors to draw on a fading, photo-sensitive panel, etc. Mostly for fun, but also to support real innovation in HCI. :)
All of that involved consciously causing physiological changes in your body for the computer to sense. Suppose this was a thought interface for an aircraft pilot. Physically they're already fully engaged in a demanding activity, what we need is to open up a new channel of communication from them to the computer in addition to their physical input. If they're focusing on twitching or changing the resistivity in their fingertips, that's going to interfere with their using their fingers to manipulate controls.
There are other similar situations, such as people in a coma that can't cause physiological changes. The point is to offload as much processing on to the computer, to make the process of communication as simple and direct for us as possible. Training ourselves to jump through mental and physiological hoops is the opposite of the point.
42 comments
[ 3.0 ms ] story [ 103 ms ] thread30-40 WPM was the target for high school computer typing courses in the 90s (my experience, may have been a similar target earlier and I don't know the target for courses targeting typewriters), and this was a time when computers weren't ubiquitous like today. It's very feasible to hit that target in a few months especially if you learn proper hand placement (hunt & peck, two finger typists, will have a longer ramp up time but can often still hit those speed targets).
https://www.tabnine.com/
> Studies compiled by Amundson (1995) show that copying rates using handwriting at the 1st grade level are about 5 words per minute (WPM) on average, but by the end of elementary school at the 5th and 6th grade level are about 10 to 12 WPM.
Source https://www.montgomeryschoolsmd.org/departments/hiat/resourc...
‘Words’ in WPM is standardized to 5 characters
https://en.wikipedia.org/wiki/Words_per_minute
Zylog Z80: 1
Intel 8086: 2
PowerPC: 4
Apple M1: 8
IBM Selectric: 5
https://news.ycombinator.com/item?id=27134049
I'd like those interfaces to be less invasive. Maybe one day they are going to be less Neo and more Case.
CTRL-labs is the one behind https://tech.fb.com/inside-facebook-reality-labs-wrist-based...
Can you even imagine StarCraft or Quake matches, or better yey, some completely new games developed for brain computer interface? Man, it's VERY exciting.
Decoding hand writing accurately is an approximate thing even when it's spoon fed to the machine via a digitizer, to say nothing of interpreting the movements from neural activity. What could be accomplished if we abandoned the familiarity of writing and existing glyphs and established a novel set of learnable 'brain semaphores' (e.g. visualize this symbol, play this sound on your mental speakers) that were easy enough for the user to practice and master with feedback, and blazing unambiguous markers to the probes and recognition algorithm? It's not hard to imagine that with deliberate practice and immediate feedback, the brain can develop connections for the sole purpose of the interface that no longer even piggyback on imagining motor/sensory events. Sending a keystroke would be no different than lifting a finger.
I'm sure this isn't a novel idea, but my armchair observer prediction is that the field will start to move in the direction of purposefully designed mind-sign schemes and away from mimicry of physical movement as the technology becomes ubiquitous and less invasive.
When I discovered NLP models like Word2Vec and Thought-Vectors, which assign vectors to words or even whole sentences/concepts, intuitively, it felt like that was exactly what was happening in my mind. It might be an illusion but I do not think in sentences or words, I only think in concepts and images that appear in my mind's eye in an instant. To form a larger idea, I build a chain of concepts. I am sure that eventually probes can pick up a clear distinct vector that uniquely summarizes the entire concept of what I was thinking or visualizing.
And as you suggested we could start off with a smaller subset of semaphores or concepts or visualizations that at first are the easiest signals to pick up via probes. After practice, these signals only get crisper.
I'm still on the far side with this, I know, but whenever these BCI stories go by I am bemused. You connect a brain to a computer and then "train a neural network" on the computer side, that makes no sense. The sophisticated NN is on the brain-side, yes? With an integrated subjective UI already, yes? You do not need fancy implants or hardware to "talk" to your machine, galvanic response and (now) micro-IMUs (inertial measurement units, one on each finger) are plenty of bits-per-second for a subjective Think-to-Type experience. Use hypnosis, it's easy, to program your brain to output a byte by changing the angle of your fingertips and then twitching the thumb to "clock" the data. Or just standard stenotype, it would likely be easier, eh?
You don't need fancy hardware or NN software (and especially surgery to stick electrodes in your brain) to "talk" to your machine with your mind, you already have all the technology you need "built in" to your nervous system. It's not even difficult to use. Literally the first thing I learned to do when I was studying (self-)hypnosis was setting up a binary "yes/no" signal (a finger twitch) from my unconscious mind to facilitate communication.
Anyhow, not to go on and on about it, the bottom line is when you connect a computer to a brain please remember that the brain is the fancier, more powerful "device", okay? You will waste less time. (I am not sure why I am not able or willing to try to make something out of this. I can talk to my computer well enough that I don't bother with BCI. ANd somehow the plight of these poor folks doesn't move me enough to do something. Am I a moral cretin? I'm seriously you guys. Should I try (harder) to get something going with simple hypnotic BCI systems? HN?)
We have so much raw data coming into our brains--via sight and sound especially, but surprising amounts via other senses as well--that I see little reason to add "additional senses" or outputs when we have rather high bandwidth already. Heck, my thoughts run ahead of my speech or typing fairly often, and both of those are pretty low-bandwidth channels.
One-hand chorded keyboards[3] are perennially reinvented by someone with an Arduino and a few buttons (not to belittle that work!) but you can't just pick up a chorded keyboard and type as fast as a regular one, so they are generally treated as a novelty. But if one spent as much time learning to use a chorded keyboard as they did a regular one, I'd bet they'd be just as fast. Perhaps the reason fluent computer users show little desire for better human-computer interfaces is that they are afflicted by a generalized form of baby duck syndrome[4]. To mix metaphors in a really confusing way, interacting with a computer with a keyboard, mouse, and monitor feels like using GNU nano: shallow learning curve, but it tapers off quickly. I wish I knew the Vim of HCI: I'd be willing to practice for years to adapt to it, to compute more efficiently for the rest of my life.
[1]: https://en.wikipedia.org/wiki/Cortical_homunculus [2]: https://en.wikipedia.org/wiki/Cortical_remapping#Plasticity [3]: https://www.stavros.io/posts/keyyyyyyyys/ [4]: https://en.wikipedia.org/wiki/Imprinting_(psychology)#Baby_d...
> the Vim of HCI
Now that is an evocative metaphor... cheers.
> I'd be willing to practice for years to adapt to it
Ah, but learn a little self-hypnosis and it won't take years.
I have a new tiny business selling custom computer peripherals: things like levers, crank-wheels, knobs, sliders, etc... for input, and odds and bods for output like sand-tables with a stylus on an arm, or lasers and mirrors to draw on a fading, photo-sensitive panel, etc. Mostly for fun, but also to support real innovation in HCI. :)
There are other similar situations, such as people in a coma that can't cause physiological changes. The point is to offload as much processing on to the computer, to make the process of communication as simple and direct for us as possible. Training ourselves to jump through mental and physiological hoops is the opposite of the point.
No, you missed the point. You use hypnosis so the physiological changes are done unconsciously; subjectively you just "think to type".