To reiterate: Human behaviors, including motor function, perception, and cognition, operate at a speed limit of 10 bit/s
I believe that the authors made the mistake of omitting the word throughput here. Just because we can’t memorize numbers or type faster than 10 bits/s doesn’t mean that cognition cannot go much faster. If someone told me to imagine a car engine, I would see it there in my mind’s eye, with a rotating crank shaft, combustion chambers, moving pistons, et cetera. Instantly retrieved. The three dimensional geometry alone is vastly more than 10 bit, let alone the animation, and it happened in the blink of an eye. The article addresses this as “subjective inflation”, even though I cannot think of a single way to compress such an imagined scene into 10 bits. Similarly, if someone told me to imagine an elephant, my mind would play elephant trumpeting, a second of which is much larger than 10 bits in any audio codec I am aware of. These few external words caused a much higher bandwidth process to unfold within the mind. There is absolutely no way that the average human mind only operates at 10 bit/s internally. I believe that the bottlenecks are learning and producing, i.e. putting things into your brain and transforming thoughts back into language and motion. But when the entire calculus happens within the mind, it can go much faster, kind of like when the program is already loaded into RAM and we can begin processing immediately, not having to wait for peripheral buses.
The fact that the inner brain isn’t slow at all answers the question “Why do we live at 10 bits/s?” - we don’t. An example: You read a theorem. The theorem is two sentences. A couple of bytes. You produce a proof. Maybe half a page. Again, a couple of bytes. You operate at 10 bit/s. But what happened in your mind between the operation of reading and writing, that is where the high bandwidth computation takes place, and that is where vastly higher bandwidths are in play. The same could be said about AI like ChatGPT: It only produces a couple of tokens per second, so it’s only 10 bits/s - simply ignoring the billions of floating point operations performed on gigabytes of parameters in the background.
The authors also lament the lack of parallelism in the inner brain. I believe that problem solving is highly parallelized, it’s just that the results are then put into a single higher level train of thought, thus giving the appearance of only being able to do “one thing at a time”.
Based on the research reviewed here regarding the rate of human cognition, we predict that Musk’s brain will communicate with the computer at about 10 bits/s. Instead of the bundle of Neuralink electrodes, Musk could just use
a telephone, whose data rate has been designed to match human language, which in turn is matched to the speed of
perception and cognition.
Despite the aforementioned issue with the paper, they may still be correct about this. Just like inspecting the parameters of an inner layer of the neural net of an AI is almost completely inscrutable, it might be the case that human thoughts are inscrutable in places where they are high bandwidth. However, it would be very interesting to interconnect the high bandwidth parts of two human brains (or to an inner layer of a neural net) and see what that would yield.
This article is more or less a summary of research going back decades, so don’t make the mistake of thinking that, just by reading it and thinking for a few minutes, you can come up with novel objections or counterexamples.
I’ve done some reading on this in the past (Tor Norretranders’ The User Illusion is a good book, though it strays into borderline quantum-woo) and my understanding is that the 10 bits/sec applies to conscious processing. Recalling a memory isn’t part of consciousness, it’s like a temporary pipe from a memory area to a sensory area. If you were to apply cognition to a memory — like counting the number of times the word “you” occurs in “I Wanna Hold Your Hand” — you’d be subject to the same bitrate as real sensory input.
The way I like to think of this is that the human brain is an immensely fast parallel computer that’s running a very complex, very slow emulator (consciousness.) The emulator isn’t really aware of the host platform, so it thinks it’s in charge and is doing all the work, when really it’s mostly an observer. (There’s another long line of research showing the mind lags our behaviors by 100ms or so, and that the brain and body initiate actions, even high level ones, before we consciously “decide” to do them.)
This article is more or less a summary of research going back decades
And that is where, at least in my understanding, the error happens: The cited material talks about input or output, i.e. throughput. Extending it to all of cognition goes beyond what the source material says.
so don’t make the mistake of thinking that, just by reading it and thinking for a few minutes, you can come up with novel objections or counterexamples.
Don’t make the mistake of assuming I’ve spent a few minutes ;-)
Recalling a memory isn’t part of consciousness
Of consciousness or of cognition? It’s definitely conscious. The definitions I am aware of squarely place memory in the basket of “cognition”. Some would even consider emotions part of cognition. Of course, if you exclude all high-bandwidth parts of our mind from cognition, then by definition cognition is “unbearably slow” for everyone.
(There’s another long line of research showing the mind lags our behaviors by 100ms or so, and that the brain and body initiate actions, even high level ones, before we consciously “decide” to do them.)
Just because we can’t … type faster than 10 bits/s
An ASCII character is 7 bits, so not being able to type faster than 10 bits/s means not being able to type more than 1.5 characters per second. The average person can type 40 WPM which is 200 characters/minute which is 3.3 characters per second. A trained court stenographer must be able to type 225 WPM of testimony, and some court stenographers can type 375 WPM, or 6.2 words/sec, which I’m sure is faster than 10 bits/sec, even taking the specialized design of steno keyboards into account. There are reports of people achieving 300 WPM on a computer keyboard, which is 25 characters per second.
What’s the fastest speed at which the average person can comprehend spoken speech? Different sources set the limit at between 300 and 400 WPM (5-6.7 words/sec). So we can comprehend spoken speech, and then type it on a keyboard, at speeds well over 10 bits/s.
The human nervous system is certainly capable of producing output (through the motor neurons) of far more than 10 bits/sec. So whatever limit the research paper is talking about is not this.
An ASCII character is 7 bits, so not being able to type faster than 10 bits/s means not being able to type more than 1.5 characters per second.
The paper addresses this in detail. When we type, we only use a small subset of ASCII in a highly structured manner. This reduces entropy significantly compared to typing characters completely at random.
it might be the case that human thoughts are inscrutable in places where they are high bandwidth.
That is precisely the point the authors make throughout, and that contradicts what you wrote earlier. This bit throughput figure is a black box estimate over observables, says nothing about internal representations.
As a former neuroscientist, this paper is meh, to me.
They talk about a “paradox” of slow output relative to input, but it’s not such a paradox at all if you assume much of the brain is dedicated to activity other than attempting to increase behavioral output rates.
Some random comments:
In fact, the 10 bits/s are needed only in worst-case situations, and most of the time our
environment changes at a much more leisurely pace. This contributes to the common perception among teenagers that “reality is broken”, leading them to seek solace in fast-paced video games (see Appendix A). Previous generations are no exceptions – they instead sought out the thrills of high-speed sports like skiing or mountain biking.
Uh, ok. I checked Appendix A, and it says nothing about reality being broken, or why people play fast video games. Instead, it talks about the bit rates of Tetris and Starcraft players.
… we could engage in any one of the 2^10 possible actions or thoughts in the next second …
At first I thought this was a typo, since it assumes a 10 bit/s output rate entails no more than 2^10 = 1024 possible actions, but apparently they wrote this in earnest.
On the subject of redundancy, again there is precious little evidence for duplication of essentially identical neurons
This glosses over the whole history of fruitless hunts for the “grandmother cell” (the one cell responsible for recalling your nana), and research on brain injuries, going back to Lashley. Its absence is either deliberate or odd.
So a large part of our paradox resolves to this question: Why is cognition restricted to one task at a time, rather than pursuing many strands – potentially thousands to millions – all in parallel?
A question which has been asked for decades, has thousands of papers written on it, multiple answers, and touches on the biggest unanswered mystery in psych/neuro: the nature of consciousness and the unconscious/conscious distinction.
…I’m going to stop here.
I get needing a strong viewpoint for your argument, but that shouldn’t come at the cost of advancing tenuous positions.
The authors primarily seem to have a background in low-level neuronal recordings, and there’s very cool things you can learn at that level, but it seems like they’re missing relevant context from higher, cognitive levels. A lot of the questions and “paradoxes” they pose have been addressed elsewhere in the literature, so their omission, even if it was just to acknowledge them, makes it feel half-baked. Obviously, you can’t address everything due to space constraints, but I don’t think that’s the driving factor, based on what they wrote.
Uh, ok. I checked Appendix A, and it says nothing about reality being broken, or why people play fast video games. Instead, it talks about the bit rates of Tetris and Starcraft players.
It makes sense to me. The preceding text talks about (young) people desiring intense experiences with fast decision-making; video games are obviously an example of this. Appendix A merely quantifies the bit-rate of some games.
The preceding text not only talks about it, but assumes it’s true. What I’m dubious about is that assertion that because young people think “reality is broken”, it leads them to seek out fast-paced video games. Their references don’t seem to support that, and there’s no shortage of slower-paced video games out there that are also popular.
Since they referenced Appendix A after the assertion, I wanted to see if it supports them, but it doesn’t.
I also checked the “reality is broken” quote, and it comes from a pop science book of the same name, Reality is Broken: Why Games Make Us Better and How They Can Change the World. Afaict, though, it’s about the social benefits (and detriments) of gaming, based on the ToC and the summaries.
If there’s something relevant in the Reality Is Broken book, they should have cited chapter/section/paragraph, but I suspect there isn’t. Given what I read, the book’s probably agnostic about game speed over game sociability. Playing a slower turn-based game probably yields many of the same benefits.
Also, it’s just not great to cite a pop science book in a paper; if the book referred to relevant original research, they should have cited the original papers, instead.
By itself, this wouldn’t be a deal-killer, but I ran into enough similar issues in the paper to consider it part of a pattern.
From discussions about this elsewhere, I learned that the meaning of “bit” in this paper, and in the cognitive-science world in general, is different than the usual meaning in the context of computers. The meaning here is more like, “the outcome of a process with two possible results where the probability of each result is around 50%,” rather than, “a single digit in a base-2 number.”
Replace the word “bit” with the word “decision” and you might end up with a clearer intuition of what they’re saying. Or maybe think of it like, the brain takes in a lot of information and can resolve roughly 10 questions per second about the ambiguities in that information stream.
I got the impression that for the authors of this paper, these two things are exactly the same. They use the word “bit” mostly in the information theoretical way, using the classic base-2-number definition and applying the base-2-logarithm, e.g.
Because the number of possible permutations of the cube is 4.3 × 1016 ≈ 2^65, the information rate during the perception phase was ∼ 65 bits / 5.5s ≈ 11.8 bits/s.
They make no distinction between the “decision” bit and a bit as used in computer science:
A much more demanding cognitive task is image recognition. A neural network model called AlexNet was trained to
classify about 1 million photographs into 1000 object categories. That means extracting 10 bits from an image
with 1.2 million bits.
Both cogsci and CS get the concept “bit” from information theory. The paper is talking about what in CS we usually qualify as “bits of entropy”, when talking about data compression or passwords.
I believe that the authors made the mistake of omitting the word throughput here. Just because we can’t memorize numbers or type faster than 10 bits/s doesn’t mean that cognition cannot go much faster. If someone told me to imagine a car engine, I would see it there in my mind’s eye, with a rotating crank shaft, combustion chambers, moving pistons, et cetera. Instantly retrieved. The three dimensional geometry alone is vastly more than 10 bit, let alone the animation, and it happened in the blink of an eye. The article addresses this as “subjective inflation”, even though I cannot think of a single way to compress such an imagined scene into 10 bits. Similarly, if someone told me to imagine an elephant, my mind would play elephant trumpeting, a second of which is much larger than 10 bits in any audio codec I am aware of. These few external words caused a much higher bandwidth process to unfold within the mind. There is absolutely no way that the average human mind only operates at 10 bit/s internally. I believe that the bottlenecks are learning and producing, i.e. putting things into your brain and transforming thoughts back into language and motion. But when the entire calculus happens within the mind, it can go much faster, kind of like when the program is already loaded into RAM and we can begin processing immediately, not having to wait for peripheral buses.
The fact that the inner brain isn’t slow at all answers the question “Why do we live at 10 bits/s?” - we don’t. An example: You read a theorem. The theorem is two sentences. A couple of bytes. You produce a proof. Maybe half a page. Again, a couple of bytes. You operate at 10 bit/s. But what happened in your mind between the operation of reading and writing, that is where the high bandwidth computation takes place, and that is where vastly higher bandwidths are in play. The same could be said about AI like ChatGPT: It only produces a couple of tokens per second, so it’s only 10 bits/s - simply ignoring the billions of floating point operations performed on gigabytes of parameters in the background.
The authors also lament the lack of parallelism in the inner brain. I believe that problem solving is highly parallelized, it’s just that the results are then put into a single higher level train of thought, thus giving the appearance of only being able to do “one thing at a time”.
Despite the aforementioned issue with the paper, they may still be correct about this. Just like inspecting the parameters of an inner layer of the neural net of an AI is almost completely inscrutable, it might be the case that human thoughts are inscrutable in places where they are high bandwidth. However, it would be very interesting to interconnect the high bandwidth parts of two human brains (or to an inner layer of a neural net) and see what that would yield.
This article is more or less a summary of research going back decades, so don’t make the mistake of thinking that, just by reading it and thinking for a few minutes, you can come up with novel objections or counterexamples.
I’ve done some reading on this in the past (Tor Norretranders’ The User Illusion is a good book, though it strays into borderline quantum-woo) and my understanding is that the 10 bits/sec applies to conscious processing. Recalling a memory isn’t part of consciousness, it’s like a temporary pipe from a memory area to a sensory area. If you were to apply cognition to a memory — like counting the number of times the word “you” occurs in “I Wanna Hold Your Hand” — you’d be subject to the same bitrate as real sensory input.
The way I like to think of this is that the human brain is an immensely fast parallel computer that’s running a very complex, very slow emulator (consciousness.) The emulator isn’t really aware of the host platform, so it thinks it’s in charge and is doing all the work, when really it’s mostly an observer. (There’s another long line of research showing the mind lags our behaviors by 100ms or so, and that the brain and body initiate actions, even high level ones, before we consciously “decide” to do them.)
And that is where, at least in my understanding, the error happens: The cited material talks about input or output, i.e. throughput. Extending it to all of cognition goes beyond what the source material says.
Don’t make the mistake of assuming I’ve spent a few minutes ;-)
Of consciousness or of cognition? It’s definitely conscious. The definitions I am aware of squarely place memory in the basket of “cognition”. Some would even consider emotions part of cognition. Of course, if you exclude all high-bandwidth parts of our mind from cognition, then by definition cognition is “unbearably slow” for everyone.
IIRC this has been debunked: https://www.theatlantic.com/health/archive/2019/09/free-will-bereitschaftspotential/597736/
Thank you! That’s really interesting. Time to readjust my world-view :)
An ASCII character is 7 bits, so not being able to type faster than 10 bits/s means not being able to type more than 1.5 characters per second. The average person can type 40 WPM which is 200 characters/minute which is 3.3 characters per second. A trained court stenographer must be able to type 225 WPM of testimony, and some court stenographers can type 375 WPM, or 6.2 words/sec, which I’m sure is faster than 10 bits/sec, even taking the specialized design of steno keyboards into account. There are reports of people achieving 300 WPM on a computer keyboard, which is 25 characters per second.
What’s the fastest speed at which the average person can comprehend spoken speech? Different sources set the limit at between 300 and 400 WPM (5-6.7 words/sec). So we can comprehend spoken speech, and then type it on a keyboard, at speeds well over 10 bits/s.
The human nervous system is certainly capable of producing output (through the motor neurons) of far more than 10 bits/sec. So whatever limit the research paper is talking about is not this.
The paper addresses this in detail. When we type, we only use a small subset of ASCII in a highly structured manner. This reduces entropy significantly compared to typing characters completely at random.
That is precisely the point the authors make throughout, and that contradicts what you wrote earlier. This bit throughput figure is a black box estimate over observables, says nothing about internal representations.
As a former neuroscientist, this paper is meh, to me.
They talk about a “paradox” of slow output relative to input, but it’s not such a paradox at all if you assume much of the brain is dedicated to activity other than attempting to increase behavioral output rates.
Some random comments:
Uh, ok. I checked Appendix A, and it says nothing about reality being broken, or why people play fast video games. Instead, it talks about the bit rates of Tetris and Starcraft players.
At first I thought this was a typo, since it assumes a 10 bit/s output rate entails no more than 2^10 = 1024 possible actions, but apparently they wrote this in earnest.
This glosses over the whole history of fruitless hunts for the “grandmother cell” (the one cell responsible for recalling your nana), and research on brain injuries, going back to Lashley. Its absence is either deliberate or odd.
A question which has been asked for decades, has thousands of papers written on it, multiple answers, and touches on the biggest unanswered mystery in psych/neuro: the nature of consciousness and the unconscious/conscious distinction.
…I’m going to stop here.
I get needing a strong viewpoint for your argument, but that shouldn’t come at the cost of advancing tenuous positions.
The authors primarily seem to have a background in low-level neuronal recordings, and there’s very cool things you can learn at that level, but it seems like they’re missing relevant context from higher, cognitive levels. A lot of the questions and “paradoxes” they pose have been addressed elsewhere in the literature, so their omission, even if it was just to acknowledge them, makes it feel half-baked. Obviously, you can’t address everything due to space constraints, but I don’t think that’s the driving factor, based on what they wrote.
It makes sense to me. The preceding text talks about (young) people desiring intense experiences with fast decision-making; video games are obviously an example of this. Appendix A merely quantifies the bit-rate of some games.
The preceding text not only talks about it, but assumes it’s true. What I’m dubious about is that assertion that because young people think “reality is broken”, it leads them to seek out fast-paced video games. Their references don’t seem to support that, and there’s no shortage of slower-paced video games out there that are also popular.
Since they referenced Appendix A after the assertion, I wanted to see if it supports them, but it doesn’t.
I also checked the “reality is broken” quote, and it comes from a pop science book of the same name, Reality is Broken: Why Games Make Us Better and How They Can Change the World. Afaict, though, it’s about the social benefits (and detriments) of gaming, based on the ToC and the summaries.
If there’s something relevant in the Reality Is Broken book, they should have cited chapter/section/paragraph, but I suspect there isn’t. Given what I read, the book’s probably agnostic about game speed over game sociability. Playing a slower turn-based game probably yields many of the same benefits.
Also, it’s just not great to cite a pop science book in a paper; if the book referred to relevant original research, they should have cited the original papers, instead.
By itself, this wouldn’t be a deal-killer, but I ran into enough similar issues in the paper to consider it part of a pattern.
From discussions about this elsewhere, I learned that the meaning of “bit” in this paper, and in the cognitive-science world in general, is different than the usual meaning in the context of computers. The meaning here is more like, “the outcome of a process with two possible results where the probability of each result is around 50%,” rather than, “a single digit in a base-2 number.”
Replace the word “bit” with the word “decision” and you might end up with a clearer intuition of what they’re saying. Or maybe think of it like, the brain takes in a lot of information and can resolve roughly 10 questions per second about the ambiguities in that information stream.
I got the impression that for the authors of this paper, these two things are exactly the same. They use the word “bit” mostly in the information theoretical way, using the classic base-2-number definition and applying the base-2-logarithm, e.g.
They make no distinction between the “decision” bit and a bit as used in computer science:
Both cogsci and CS get the concept “bit” from information theory. The paper is talking about what in CS we usually qualify as “bits of entropy”, when talking about data compression or passwords.