Slashdot Mirror
Ray Kurzweil Does Not Understand the Brain
jamie writes "There he goes again, making up nonsense and making ridiculous claims that have no relationship to reality. Ray Kurzweil must be able to spin out a good line of bafflegab, because he seems to have the tech media convinced..."
The singularity is to nerds what the rapture is to fundamentalist protestant wackjobs....
Would be nice if the summary even hinted at what the ridiculous claim actually WAS...
Namely, that we'll be able to reverse engineer the human brain in the next 10 years.
...he must be right! He used math, and everything! I'm a little shocked that Kurzweil equates blueprints with the functioning organ. I am not shocked, however, that the tech media latched onto this--at first blush it sounds so *reasonable.*
Kurtzweil is just the latest in a long line. How do you think publications like Fast Company and Wired get written?
And does anybody remember JonKatz?
His latest claim is that we'll be able to reverse engineer the human brain within a decade
Amateur. I could put something together to simulate the human brain in about 8 months.
(Plus another 3 minutes at the start)
Summation 2
Sejnowski says he agrees with Kurzweil's assessment that about a million lines of code may be enough to simulate the human brain.
Here's how that math works, Kurzweil explains: The design of the brain is in the genome. The human genome has three billion base pairs or six billion bits, which is about 800 million bytes before compression, he says. Eliminating redundancies and applying loss-less compression, that information can be compressed into about 50 million bytes, according to Kurzweil.
About half of that is the brain, which comes down to 25 million bytes, or a million lines of code.
Idiot. The design of the brain is encoded in the genome in the same way that the design of a 4KiB program is encoded in its load module: useful for running the program on its original hardware.
But then you have architectural issues. That 4KiB of information does not run unless it's supported by a complex operating system, which itself is supported by complex logic in an ISA and memory managment architecture backing it up. And all that is implemented on a specific design in a specific physics model.
Translating that program to SPARC takes work, and it comes out roughly the same size. Translating that program to a progression of chemical reactions produces something vastly different, especially since you need a new middle ware (chemical environment) running on top of different physics (chemistry).
Translating a physical architectural design from chemistry to computer logic on top a given ISA is the same problem. You now have odd issues that are messy, and then the program running on the brain needs to be built again. That program is even more complex and less known.
Support my political activism on Patreon.
Ray Kurzweil is yet another computer programmer blathering on about things that he has no understanding on.
Get Kurzweil a slashdot account, stat!
Would be nice if the summary even hinted at what the ridiculous claim actually WAS... Namely, that we'll be able to reverse engineer the human brain in the next 10 years.
It's a little more complicated than that. You see, the article actually breaks down the logic behind that statement and points out how poor it is. Here's the initial part of Kurzweil's argument:
Sejnowski says he agrees with Kurzweil's assessment that about a million lines of code may be enough to simulate the human brain.
Here's how that math works, Kurzweil explains: The design of the brain is in the genome. The human genome has three billion base pairs or six billion bits, which is about 800 million bytes before compression, he says. Eliminating redundancies and applying loss-less compression, that information can be compressed into about 50 million bytes, according to Kurzweil.
About half of that is the brain, which comes down to 25 million bytes, or a million lines of code.
I have only taken high school biology but I know that the genome doesn't magically become the brain. It goes through a very complex process to amino acids which fold into proteins which in turn make cells that in turn make tissues that in turn comprise the human brain. To say we fully understand this transformation entirely is a complete and utter falsity as demonstrated by our novice understanding of the twisted beta amyloid protein that we think leads to Alzheimer's. How amino acids turn into which proteins I believe is largely an unsolved search problem that we don't understand (hence efforts like Folding@Home). And he claims that in ten years not only will we understand this process but we will ... reverse engineer it?
The man is insane. I've posted about this same biologist criticizing him before and it looks like P.Z. Myers just decided to take some extra time to point out how imprudent Kurzweil's statements are becoming. Kurzweil will show you tiny pieces of the puzzle that support his wild conclusions and leave you in the dark about the full picture and pieces that directly contradict his statements. This is a dangerous and deceptive practice that -- despite my respect for Kurzweil's work in other fields -- is rapidly turning me off to him and his 'singularity.' He's becoming more Colonel Kurtz than Computer Kurzweil.
My work here is dung.
Let's see. On another recent article it was stated that the average car has several million lines of code running in it. I haven't come across a sentient Prius yet.
And there's that pesky parallel processing the brain does. I don't think that a rack full of Nvidia Tesla cards can approach the average two year old's parallel processing capability.
I agree, Kurzweil is smoking something and not sharing.
No not really.
A computer is a fixed system. If you tell it to do A (via software), you know you will get B, based upon knowledge of how the circuits are hardwired. The same can not be said of the human brain, because it has the ability to change its hardware (via growing new connections between neurons).
"I disapprove of what you say, but I will defend to the death your right to say it." - historian Evelyn Beatrice Hall
I haven't read the article yet, but Ray Kurtzweil is a technology speculator - like a sci-fi writer except that he doesn't make up a story to go with his ideas and tries harder to convince people they're actually going to happen. He wrote "The age of intelligent machines" and "The age of spiritual machines" where he takes a hard AI stance that computer thought can become indistinguishable from human thought. He is also a proponent of technological singularity.
Generally his ideas aren't taken very seriously by academia in Computer Science, or at least that has been my experience. The philosophy department at my university sometimes enjoyed going over his ideas; but the philosophy department at my university was very fond of pseudoscience.
Sejnowski says he agrees with Kurzweil's assessment that about a million lines of code may be enough to simulate the human brain.
Kurzweil explains: The design of the brain is in the genome. The human genome has three billion base pairs or six billion bits, which is about 800 million bytes before compression, he says.Eliminating redundancies and applying loss-less compression, that information can be compressed into about 50 million bytes, according to Kurzweil.
Dude, the equations of quantum mechanics can be written on one page. General Relativity can be written on a second page. What more do you need ? Clearly, a few hundred lines of code (and a few do loops) should be enough to simulate the entire universe, brains and all.
Glad we cleared that up. All you physicists and astronomers can go home now and work on your resumes.
You're being conveniently trite here, though. That's not a good counter-argument. This particular biologist seems to have a pretty good grasp on the fundamental problem with Kurzweil's argument, and that problem is: Kurzweil confuses the purpose of the genome. It is not "the program"! Myers contends that, really, it's more like data. To me, this sounds like a classic Von Neumann architecture: it's bit of both, depending on your context. In any case, Kurzweil completely misses out on the fact (and he would know this if he had followed *anything* in genomics over the last 15 years) that the genome, as encoded in DNA, is only a small part of what makes a cell express and function in a particular way. A nice introduction to the epigenome was in this NOVA documentary.
To simplify it so a computer science ignorant biologist with a tendency to inane rants can possibly get it, you don't need to simulate electrons in a semi-conductive material at specific temperatures in order to build a complete working emulator for an old computer.
"I love my job, but I hate talking to people like you" (Freddie Mercury)
...except, maybe, Pinky.
Guaranteed! This comment 100% Anthrax free!
He actually has one.. And he's a dick, too.
Obviously you've never heard of FPGA: http://en.wikipedia.org/wiki/Field-programmable_gate_array While you can't add new connections in the strictest sense, you could could conceivably create a chip with a whole bunch of generic unused hardware and in the rest of the hardware program an algorithm that allows new connections to be made with that raw material.
Visualize Whirled Peas
Scanning down the comment list, it looks like every (+2 or more) comment has read the article and is quoting from it -- what has happened to the slashdot I knew and loved?
I mod down anyone who says "I will be modded down for this", regardless of the rest of their comment
You are assuming 1 neuron = 1 "byte" of data.
It's much more complex than that. We are barely starting to understanding it now.
I agree with you, though, if you are implying that the brain is a physical entity with a physical size and physical limits. We just don't quite know yet what those limits are.
"Whenever people agree with me I always feel I must be wrong." (Oscar Wilde)
Which account is it?
PS: Go fuck yourself.
if you rtfa you'll see that the million lines of code only gives you the proteins that make up the brain - i.e., it gives you a parts list and a delivery schedule, not a set of assembly intstructions. The genome doesn't give you how the proteins interact, in usually complex ways (i.e., three or more proteins interacting simultaneously), in billions of cells in parallel, over the course of 9 months to give us an infant brain (even leaving aside the tremendous amount of development that takes place in the brain during childhood).
As the author of tfa writes: To simplify it so a computer science guy can get it, Kurzweil has everything completely wrong. The genome is not the program; it's the data.
IOW, the program is the developing organism itself, the complex protein interactions and it's (uterine) environment none of which are encoded in the genome. The organism uses the data encoded in the genome to produce proteins which interact with each other and the organism and its environment to grow cells which eventually form a brain.
The mistake in Kurzweil's thinking is the typical mistake engineers make when dealing with biology; the enviroments into which engineers place their designs do not typically spontaneously cooperate in the construction of the engineer's design. When an engineer designs a circuit board, his lab bench doesn't spontaneously start soldering connections and adding components for him and automatically complete parts of the design
without his explicit instructions. But the organism does precisely this with proteins syntesised from the genome.
As a result, the genome alone cannot possibly tell you how to "make" an organism, because the genome only tells you the parts list and delivery schedule for the organism, not the assembly instructions. The assembly instructions are not explicit anywhere in the system; the assembly instructions are implicit in the combination of the complex behavior of the cells of the developing organism, the uterine environment and the very complex ways the proteins sythensized from the genome interact.
In order to extract the actuall assembly instructions we'd need a full blown molecular biology simulator that could correctly simulate:
1. protein folding (still unsolved)
2. comlex multi-protein interaction (still unsolved)
3. simultaneous behavior and development, (i.e., in parallel) of billions of living cells each undergoing trillions of chemical reactions per second (computationally prohibitive)
IOW, it's not going to happen in the next 10 years.
Yes. Well done. Did you try reading the article that you are criticising because it rips your point apart fairly easily. The thing about an upper limit is that it should be at least as large as the thing that you are estimating. The article shows quite conclusively that Kurzweil's "upper limit" is far too small because he knows nothing about brains and pulled some numbers out of his arse.
That "tangent" that Myers went off of was a reasonable argument for why the amount of information described is not sufficient to simulate a brain. Not least because it is a highly compressed description of a process that builds a brain. It is not a description of a brain itself. Furthermore to use that description to build an actual model of the brain you need to understand all of the biological processes that are relevant in executing that construction code, and the environment that they run in.
Oh the irony, it's burning my eyes. You're defending somebody who was caught babbling about something they don't understand by repeating the trick. Well done you.
Slashdot: where don knuth is an idiot because he cant grasp the awesome power of php
I'm not sure what it is about his claims that are supposed to be so ludicrous. For example, a million lines of code seems at least plausible, as long as we bear in mind the following:
1. We're not trying to mimic the brain at the protein level, rather at the broader, inter-neuron level (and whatever complex intra neuron behaviour we discover).
2. The million lines of code don't need to encompass the capacity of the brain, just its general neural architecture and adaption rules - there will no doubt be many gigabytes (terabytes?) of working memory, which would actually store the neural connections and whatever parameters they may have.
To be honest, the authors of this article seem to be rather too cocksure in dismissing all this. Even the apparent agreement of Terry Sejnowski (co-inventor of the boltzmann machine) doesn't give them pause. I'm not that familiar with Kurzweil's predictions, but this seems fairly reasonable to me.
There is a google tech talk by Geoff Hinton on restricted boltzmann machines, (a sort of stochastic neural network) that's well worth a watch, for those that are interested. They are considered biologically plausible, and he seems mostly to apply them to machine vision tasks.
Ray's documentary about the Singularity has been touring the national film festivals along with Ray. I saw it June. Its begins as a dull-talking heads piece about the current state of A.I. Many of the Big Name A.I. Scientists are interviewed. Then it transitions into a crime-drama story about the legal rights of A.I.s. That part was more interesting, since it had a story. The film is full of special effects to advance the story. Although I know most of the film to be factual, I suspect it will look like a scifi movie to the average audience member. I think Ray is seeking looking at broader distribution on cable television or arts theaters after the festival run.
Ray was interesting in person during a film-makers Q&A. He reminded me of Woody Allen, but more confident and intelligent. He was graduated from M.I.T. about decade before myself. I personally believe in the Singularity, but more likely in centuries rather than decades.
Haha, see what I mean?
There are currently four academic disciplines working on the reverse engineering of a human mind. Linguistics, psychology, computer science, and philosophy. You can count neurology too if you want to start talking about the *actual* brain. Several tens of thousands of individuals are directly and indirectly working on this problem. We've come a long way in the last few decades. Unfortunately, we have a pretty long way to go. For the moment we lack a model which accurately describes how mental processes work. There isn't even a consensus on how the processing is done.
"modeling the brain" is not even really the hard part. One only needs sufficient computing power to model what they *think* is going on logically (there isn't even a consensus here). The trick is modeling the mind. We are very, very far away from that.
A fun number to throw around is how many synaptic connections are present in the brain. Synaptic connections are widely believed to be the best indicator of overall memory storage and processing speed (to an extent). There are about 10 to the 15th (Peta I believe?) synaptic connection in a normal human brain. A significant number of these are active at any given time. In other words, the brain is performing a HUGE number of "calculations" simultaneously at all times. Modeling just the hardware is obviously not easy... modeling the software is currently not possible. I doubt it will be in the next 50 years.
For a good read on what many cognitive scientists think is going on, though it is clearly not an accurate model but rather a best guess, go read up on "connectionism".
Obviously, by your logic, a free market economy is impossible, Our economy is too complex to have evolved on its own. In fact, it is far more complex, with far more different parts, than a human being. It must have had a creator. If most any part of the economy, like the steel industry, say, were removed, the economy would not function. How did the economy function before there was a steel industry? Obviously, it couldn't, and therefore we have demonstrated irreducible specificated complexification or something.
All this free market talk is obvious bullshit, and we actually DO have a centrally planned economy because it is impossible for something so complex to have evolved without a central planner.
- None can love freedom heartily, but good men; the rest love not freedom, but license. -- John Milton
First off, Ray Kurzweil doesn't want to die. That's a preoccupation that a lot of people have (including one of his critics, Rudy Rucker, who has written whole books hoping to find immortality in the fourth dimension), and it leads them to some pretty fantastic conjectures from time to time. It's not necessarily a bad thing, as long as you keep the proverbial grain of salt handy. Modern chemistry and its not insignificant contributions to our vastly expanded lifespans arose from the alchemical search for immortality. Alchemy was bullshit, of course, but the incidental discoveries of alchemists on the way to their illusory elixir of life paved the way for the real science to follow and build upon after it had ejected the dross.
And secondly, I don't think it's entirely implausible that we can eventually design hardware and software that will match and exceed the performance of the human brain. Our brains, after all, are the end product of evolution, and like pretty much every other part of our bodies, an accumulation of kludges that were just good enough to get passed to the next generation (or not bad enough not to get passed on). It's also implemented using hardware so unreliable that it wouldn't function at all if it wasn't constantly repairing itself, and even then, no matter how well you treat it, it irreparably craps out after about 75 years. And it still doesn't work all that well -- ever seen the long chain of train wrecks that is the history of human civilization? We might be able to engineer something that works a lot better. Granted, it's not going to be by deriving simulated human brains from a copy of the human genome. More likely, it will be very much unlike the way biological brains work.
The fundamental problem, which I think smart and optimistic guys like Ray Kurzweil are particularly prone to forgetting, is that it may not be possible for a mind to understand a mind of equal complexity, i.e., humans may lack the necessary intelligence to duplicate their own intelligence. That will force us back on genetic algorithms to evolve AI, leading to an end product that will likely be just as badly undesigned as natural brains. Worse, it will do little to advance our understanding of how minds work: if we can't reverse-engineer our own brains, we probably won't be able to reverse-engineer even more sophisticated artificial minds, nor will they be able to reverse-engineer themselves. (We can hope that they could reverse-engineer us, and then explain it to us in terms we can understand, if such terms exist, but that takes us so far out on a conjectural limb that I can see Ray Kurzweil from here.)
Anyway, there's room for bold conjectures. That doesn't mean that when Kurzweil completely fails to understand the way molecular biology works that we shouldn't call bullshit on it, but we shouldn't be entirely hostile to futurist speculation. By nature, most of it will be bullshit, but a lot of progress in unexpected areas has been made in the pursuit of mirages (alchemy leading to chemistry, astrology leading to astronomy), and explaining (or discovering) why a conjecture is bullshit is a beneficial exercise in and of itself.
Proud member of the Weirdo-American community.
PZ Myers threw a red herring there. What Kurzweil says is pretty reasonable, he used the total amount of information in the genome to get an upper limit estimate of the amount of library code needed to simulate a brain. I say "library" to differentiate from data, since a lot of our brain information comes from our experiences, i.e. library == instincts.
Actually he's right. The statement is pure bullshit.
Or maybe that's too much. Kurzweil just doesn't understand how Kolmorogrov complexity works.
Let's say the brain as a machine is the output of a process. How complicated is that process? The Kolmorogrov complexity of a string (or whatever) is the minimum size of the data that you have to give to a machine in order to produce the string. E.g. a string of 100 0s is simpler than a string of alternating 0s and 1s and simpler than encoding the first 100 digits of pi. Write code for each of those and you'll see the measure works (and it's actually a lower limit, but it's the closest concept...)
But the crucial point is that the size of this string depends on the kind of machine. The size of the input (program) for a Turing machine is very different than that for an actual computer.
So, yes. 800MB of code. But that's not the library code. The library that interprets that program is the egg that grows those 800MB of data into a human, together with all the laws of physics and chemistry involved in the process.
Take all the chromosomes encoding a whole human genome and drop it into a test tube of distilled water. Does it grow a brain? What if you put it into a chicken egg. What grows out? Putting those 800 MB into a computer doesn't do anything if you don't provide the equivalent of the egg. The bootstrap structure and the underlying architecture are as important as the code in understanding the whole system.
Myers is right. In order to understand the human brain directly from the genes you have to understand all chemistry that interacts with it, all the self replicating machinery provided by the mother and simulate that at a molecular level.
So the upper bound is NOT 800 MB. It's 800 MB plus the size of a codebase good enough to simulate every interaction at an atomic level plus a full 3D scan at an atomic level of the egg provided by the mother. Or simplified models of all those things, provided by the chemists and biologists out there, as Myers points out. (Plus data equivalent to a few years of training like we do with children)
Not saying that simulating the brain is necessarilly that hard, it's just that Kurzweil's pseudo-scientific measurement is just bullshit.
The human brain is composed of one hundred billion or so neurons. Looks like it's pretty much finite to me. I have ten times as many bytes of information in my hard disk.
Yet while you were typing (presumably not saving anything other than in RAM), was the content of your hard disk changing (Yes, perhaps a bit, but play along for this example)
The neurons are continuously 'remapping' in your brain. Even while some may be static, other's are making new connections in manners which we currently can't predict, or really understand why did it connect to 'this' neuron instead of 'that' neuron.
Not that the brain functions in any quantum manner, but it's one of those things that if you were to KNOW the exact mapping of neurons, the very next instant the mapping would be incorrect and very quickly become inaccurate (100 billion or so items making new connections in multiple paths)
I suppose it would be something like trying to map the water vapor droplets in a cloud. There is a finite number of droplets there too, but predicting the shape/behavior of a cloud with any precision after only a single second would be very, very difficult.
Out of modpoints but really liked a post? 1BDkF6TtmmeZ3yqXbz9yhdYVqRYnwFoXDj
A model of the human brain would need to model 10^10 neurons, each connected (not at random) to some 10,000 other neurons to produce a net of 10^14 synapses.
To understand the challenge of modelling a system this vast and complex, consider the state of research on the model organism Caenorhabditis elegans (a tiny worm). After many years of work its nervous system has been (almost) exactly mapped: it contains 302 neurons, 6393 chemical synapses, 890 gap junctions, and 1410 neuromuscular junctions. Imagine now the difficulty of reaching this level of precision in a system 10^7 times larger. Unlike the genome, we have no clues about how to automate mapping of an intact brain.
But the good news is that with this level of neuro-mapping precision we can now completely simulate the neural network ("brain") of a tiny worm, right? Right?
Wrong. Not by a long shot. We are still struggling with characterizing the behavior of this primitive neural net, and making efforts at simulating some aspects of that behavior. The 302 neuron "brain" is far beyond our abilities to simulate at present.
Starships were meant to fly, Hands up and touch the sky - Nicky Minaj