Mapping The Brain To Build Better Machines (quantamagazine.org) 110
An anonymous reader quotes a report from Quanta Magazine: An ambitious new program, funded by the federal government's intelligence arm, aims to bring artificial intelligence more in line with our own mental powers. Three teams composed of neuroscientists and computer scientists will attempt to figure out how the brain performs these feats of visual identification, then make machines that do the same. "Today's machine learning fails where humans excel," said Jacob Vogelstein, who heads the program at the Intelligence Advanced Research Projects Activity (IARPA). "We want to revolutionize machine learning by reverse engineering the algorithms and computations of the brain." By the end of the five-year IARPA project, dubbed Machine Intelligence from Cortical Networks (Microns), researchers aim to map a cubic millimeter of cortex. That tiny portion houses about 100,000 neurons, 3 to 15 million neuronal connections, or synapses, and enough neural wiring to span the width of Manhattan, were it all untangled and laid end-to-end.
If you want to avoid religious issues (Score:2)
Stay away from the right parietal lobe...
Better yet - stay away from both lobes (Score:2)
There are 7 billion people and counting on this planet - why do we need a build a poor electronic fascimile of a human when we have so many real brains here? This is nothing more than ego on the part of the researchers.
And before someone quotes the industrial revolution at me - that replaced physical strength, something that humans even compared to other animals are poor at. However we are exceptionally good at thinking (as a species , not necessaily per individual) so other than the glory of the people inv
Re: (Score:2)
We already have thermostats and controllers. They don't have to act like a human brain, and in fact, they shouldn't act like a human brain that might get distracted or confused. There is absolutely no need to make controllers more "brain like", but the meme has been put out there by Hollywood, so you get people wasting time like this. Mapping the brain is critical to understanding better how it works, but trying to make computers act more like brains is just not a sound scientific concept. Brains and compu
Re: (Score:3)
but trying to make computers act more like brains is just not a sound scientific concept.
You may not think it's useful, but there's nothing unscientific or unsound about it. It's a matter of understanding how the brain works, and throwing enough hardware at it to duplicate the essential operations.
Re: (Score:2)
You will need to make it out of nerve cells and glia then, because silicon won't cut it.
Re: (Score:3)
You will need to make it out of nerve cells and glia then, because silicon won't cut it.
Just like air planes need to be made from bone, muscle and feathers, because otherwise they won't fly ? Seriously, what's so special about nerve cells that we can't duplicate on a functional level ? And how do you know that to be true ?
Re: (Score:2)
I love it when engineers try and tell biologists how biology works. But please, be my guest and try and make a brain out of silicon. I am a neuroscientist by profession, and I and many neuroscientists outside of the cognitive neuroscience subfield say that if you want to make something that works like a brain, you will need to do it with wet-ware, not hardware. But I am not discouraging you from trying. Give it a shot and get back to us biologists in a couple decades with your results.
Why can't you do it? B
Re: (Score:3)
Because there are far too many things happening at too small a scale to even measure, let alone imitate. The unknowns in neurosicence far outnumber the knowns.
Take an intel Core i7 in a time machine, and drop it on someone's desk in the 60's. Ask him to imitate it. You'll probably get the same reply.
You can model certain brain functions with software to see what happens if you alter inputs to a neural circuit, but it will only be as good as weather predictions done in silicon.
Irrelevant. The reason we can't do good weather predictions is because weather is chaotic by nature.
Re: (Score:2)
So why hasn't it been done? And what makes you think that weather is more chaotic than neural activity? What you are saying is coming from a lack of knowledge about how the brain works. You know why? Because neuroscientists don't know why either, and they spend all day studying it. You can't imitate something that you don't understand. Understand?
Re: (Score:3)
And what makes you think that weather is more chaotic than neural activity?
Even if neural activity is chaotic, that only means we can't perform the same thing as a brain with 100% accuracy. But
Re: (Score:2)
Like I said, good luck. People have been trying for decades, and haven't gotten anywhere near emulating brain activity in silico. I am not saying people shouldn't try, They will discover all sorts of interesting things along the way, but they will not come up with anything like brain activity in a computer.
Your plane analogy was good, they don't reproduce, they don't flap their wings, they can turn on a dime in mid air with a flip of their wings, and they can't land on a tree branch. They can't build nests
Re: (Score:3)
That is not making an artificial bird, that is making a plane that flies people around. They are not even slightly the same.
The goal of a plane is to fly. The goal was never to mimic a bird. Similarly, the goal is to make a computer that can do tasks that our brain can do, but we don't have to make it run on ham sandwiches and milk.
Re: (Score:2)
People have been trying for decades, and haven't gotten anywhere near emulating brain activity in silico.
What about the the Blue Brain Project? They have simulated a complete rat neocortical column in silico, and it reacts to stimuli in the same way as a wetware one does. I don't necessarily think this is the path to human-level AI (evolution often has a lot of unused baggage, simply because a particular mutation didn't reduce fitness) but a working model of a brain (human or rat) would be incredibly beneficial in identifying root causes of neural problems and testing treatments.
Re: (Score:2)
It is a simulation. A pretty good one, but just a simulation. Cortical columns don't work in isolation, and they are microscopic in width. There are vast numbers linked in extremely complex ways in cortex, and that is just neocortex. Then there are the sensory systems, the motor systems, thalamus, hypothalamus, the entire midbrain, and then the hugely complex cerebellar cortex and deep cerebellar nuclei, plus brainstem and spinal cord, not to mention the peripheral nervous system. Then there are the vast fi
Re: (Score:3)
Like I said, good luck. People have been trying for decades, and haven't gotten anywhere near emulating brain activity in silico.
Except for the ANN that just won a Go tournament. Or the ANNs that can do face recognition, speech recognition, or many other things.
Of course, they can't do everything that a brain can do, but a baby can't run a marathon either.
Deep learning based on sigmoidal belief nets is inspired by the architecture of the brain. Autoencoders are very similar in function to the "mirroring" that occurs in the brain. Silico and vivo are not as different as you believe.
Re: (Score:2)
just like simulating a brain in a computer is not going to give you anything like a mind
If you put a computer inside a robot skull, and attach its inputs to sensors and outputs to motor control, the distinction between a real mind and a simulated one disappears. It's just a different form of mapping the same information.
Re: (Score:2)
Sure, that must be how it works. I am sure that it would be a self aware, thinking, feeling robot, because that's what I saw in the movies.
Again, good luck with that. You will need a lot of luck.
Stating something is going to work, and getting something to actually work are very different things.
Re: (Score:2)
So you think that the ANN is thinking like a person does? That would be very interesting considering we have no idea how the brain works.
Re: (Score:2)
I am sure that it would be a self aware, thinking, feeling robot
As long as your simulation is good enough, you'll get a good representation of all the things you mentioned, yes.
Stating something is going to work, and getting something to actually work are very different things
I know. But you said it wasn't going to work, so I'm only arguing the first part, not the second.
Re: (Score:2)
So you think that the ANN is thinking like a person does?
Probably not, but I also don't think that two human Go players think the same way, so I wouldn't consider it a critical point.
Re: (Score:2)
It won't work if you are trying to make a simulated brain that works like a brain using silicon. It may imitate the brain in superficial ways, but that's about it. I am sure they will make good advancements in computing, but not in artificial minds. They will just be putting imitative lipstick on a hardware simulation.
Re: (Score:2)
Two brains use slightly different neural connections based on their genetics and experience, but comparing those subtle differences to the differences between how a brain functions and how a computer functions is silly. I completely understand that non-biologists, who don't understand biological complexity, can think that making an artificial brain is quite doable. But considering we are decades, if not centuries, away from even sort of understanding how the brain works, I think you are going to have a bit
Re: (Score:2)
It won't work if you are trying to make a simulated brain that works like a brain using silicon. It may imitate the brain in superficial ways, but that's about it.
That's entirely dependent on the accuracy of the simulation, and there's no theoretical limit to that.
But considering we are decades, if not centuries, away from even sort of understanding how the brain works
I don't think we ever will, at least not in detail. I also don't think it's necessary. Nobody understands how the AlphaGo neural net works, but it can still beat the best human players in the world. The net just taught itself, first by looking at examples and then by playing against itself.
Re: (Score:3)
I completely understand that non-biologists, who don't understand biological complexity, can think that making an artificial brain is quite doable.
I remember chess grandmasters arguing in the 80's that chess computers would never be able to beat them, unless they found a way to understand how a grandmaster plays, and somehow put that in code. Botvinnik wrote a book in 1984 about how computer programs should formulate long term plans. Now, with modern computer hardware, an expert programmer could write a chess program that would beat those grandmasters, while himself not possessing any more chess knowledge than can be picked up from a beginner's book.
Re: (Score:2)
I don't see how that is creating an artificial brain, it is just creating an artificial game playing machine. I find it interesting that you think you can make an artificial version of something without knowing anything about how the original works. That is a fascinating concept.
Re: (Score:2)
Well then you should have an artificial brain completed shortly, right?
Re: (Score:2)
So you think that the ANN is thinking like a person does?
At the neuron level, yes, they basically function the same. Brains seem to use an algorithm similar to backpropagation, so "correct" responses lead to strengthened connections, and "incorrect" responses cause connections to be weakened.
That would be very interesting considering we have no idea how the brain works.
That is nonsense. We don't (yet) have a complete model for the brain, be we have far more than "no idea".
Re: (Score:2)
Brains are biological. Your brain is an organ in your body like your liver is. It requires nutrients, oxygen and a good blood supply. Brains don't use algorithms, computers do. I have been a publishing neuroscientist since the 1980s and I am unequivocally telling you that we don't know how the brain works. Prove me wrong. Get on PubMed and bring up the articles that declare that humans know how the brain works. We haven't got a clue because we don't even know why life is so different from inanimate matter.
Re: (Score:2)
You can't imitate something that you don't understand. Understand?
Yes you can. The people that developed Alpha-Go are not particularly good Go players. Yet they created a program that can easily beat its creators. Many systems display emergent properties that are were not planned by the designers.
Re: (Score:2)
"Just like air planes need to be made from bone, muscle and feathers, because otherwise they won't fly "
You might want to compare the efficiencies and aeronautic abilities of aircraft vs birds.
Re: (Score:2)
You might want to compare the efficiencies and aeronautic abilities of aircraft vs birds.
That may be interesting, but that's not really the point. I'm not arguing that aircraft and birds are equivalent, rather that they can both fly, even though very smart people once claimed that machines would never be able to do that. Now, other smart people are saying machines will never be able to think. And the only argument is from incredulity.
Re: (Score:1)
Re: (Score:2)
Even insects appear to get bored so any sufficiently complex neural net may well exhibit similar properties. If you want a mindless automaton that does the same task over and over you're better off with a programmed computer with maybe a tiny neural net for some pattern matching, not a brain simulation which is what these guys are aiming at.
Re: (Score:2)
That's easy: You just erase the used net after each computation and copy it back from the known-good state.
I don't think they want a brain simulation, skimming through. It looks more like they want a specialised neural net, but don't know how to build one. So they are mapping a chunk of brain and will try to figure out how it works, and use the knowledge thus gained as a guide for creating specialised artificial networks for visual processing.
Re: (Score:2)
Re: (Score:2)
Not true - all organic brains require sleep.
Re: (Score:2)
Humans thinking is not particularly good, it is actually quite poor compared to a reasonable ideal.
A properly developed AGI will probably be able to solve many problems that are difficult or even impossible for humans to solve.
Pure delusion (Score:5, Interesting)
This cargo-cult approach to AI is ridiculous. Decades of effort have produced absolutely no result. Oh, but this time we're way smarter and better informed, surely we'll produce something of value this time. Gimme the grant monies, plz.
Oh, but this one is worse. It's not that gigantic failure. The laughable failure they're repeating this time is far, far, older: "We want to revolutionize machine learning by reverse engineering the algorithms and computations of the brain."
Computationalism?! Seriously? Not only is that laughable, it's been laughable for ages! Don't think so? People have been born and died of old age waiting for that bit of fiction to produce any results. So far? Nothing. On top of it all, there's more than one good reason to suspect it's never going to produce any results.
Let's base one retarded idea on another retarded idea and mix in a bunch of childish thinking about the function of the brain based on zero evidence. AI breakthrough!
Re: (Score:1)
Cynicism isn't productive or warranted. The mechanisms of the brain are already the basis of the most useful machine learning algorithms. Statistical inference only goes so far and really relies on assumptions that limit application to an ever smaller subset of the real world's problems. Don't let the problems from media hype and marketing jargon delude you into ignoring the real and practical utility of this approach. It is modeling the network of that region of the brain and using our existing cognitive p
Re: (Score:2, Interesting)
Cynicism isn't productive
Neither is computationalism. It's like saying they're going to crack strong AI through phrenology.
or warranted.
Oh, I'd say it's warranted. What other reaction could a reasonable person have to this? Imagine if someone announced that they're going to make significant advances in perpetual motion thanks to phlogiston theory. That's exactly what this sounds like to anyone who isn't a Kurzweil cultist.
Here's a neat idea. Let's let go of old, long disproved, ideas and try something new. The alternative, after all, is t
Re:Pure delusion (Score:5, Insightful)
This cargo-cult approach to AI is ridiculous. Decades of effort have produced absolutely no result. Oh, but this time we're way smarter and better informed, surely we'll produce something of value this time. Gimme the grant monies, plz.
Oh, but this one is worse. It's not that gigantic failure. The laughable failure they're repeating this time is far, far, older: "We want to revolutionize machine learning by reverse engineering the algorithms and computations of the brain."
Computationalism?! Seriously? Not only is that laughable, it's been laughable for ages! Don't think so? People have been born and died of old age waiting for that bit of fiction to produce any results. So far? Nothing. On top of it all, there's more than one good reason to suspect it's never going to produce any results.
Let's base one retarded idea on another retarded idea and mix in a bunch of childish thinking about the function of the brain based on zero evidence. AI breakthrough!
What is your approach to building a strong AI then? We are waiting for your reply.
In other news, the amount of progress into AI research depends on what fronts you judge the progress and there have been numerous steps forward but none of them have resulted in C3P0 style robots because that is not the goal.
There are functionalists who largely congregate at MIT and Harvard who do not believe that studying the human brain will yield any immediate results that can be implemented in silicon
There are the Cal tech researchers who are largely behavioralists who believe studying psychology and to a limited degree physiology will point the general direction in terms of large milestones that have to be achieved to build a strong AI, using nature as a guide.
Then there are the Connectionists and this includes Francois Crick, who believe that the neural connections as a functional network will yield a base cortical algorithm which can be applied to a number of things that also include the beginnings of building a Strong AI.
There are a lot of things that have been gained by these approaches to the problem so to say as you did that nothing has been accomplished is about as wrong as wrong can get. We know for instance, that the processing "program" of the human brain uses the same functional unit that is repeated over and over and is adapted and adaptable to vision processing, audio processing, kinesthetic processing and very likely (almost certainly) everything else in terms of recognizing adapting to and predicting (prediction is a strong indicator of intelligence among other behavioral emergent patterns) patterns and sequences of patterns between inputs and outputs. If you surgically connect auditory nerves to visual cortex the visual cortex adapts to process inputs from the eardrums, if you connect optic nerves to audio cortex, audio cortex processes visual information.. with no further manipulation.. that is quite a big value for "nothing being accomplished" as you put it.. Heres the kicker of how wrong you are:
A camera device has been developed that non-invasively communicates digital information as points of pressure onto the tongue of the wearer, thereby allowing a completely blind person to navigate and "see" through the camera via the sensation on the tongue. But no, we have not accomplished anything involving the understanding of how the brain processes information and yes we just should abandon this line of research because it will not accomplish anything at all.
You are right man, Deep Blue did not beat Kasparov and no IBM's Watson did not win against champion humans on Jeopardy.. so we might as well not even try.
Sheesh! Don't even bother to answer unless you have actual points to make with cited references.. go back to putzing around in Minecraft ok?
Re: (Score:1)
What is your approach to building a strong AI then? We are waiting for your reply.
I don't have one. Of course, neither does anyone else. That said, beating on long disproved approaches isn't exactly going to get us anywhere.
Computationalism is as dead as spontaneous generation. You don't need an alternative to find out that something doesn't work, and is never going to work. You'd have us repeat the same failure over and over rather than work toward finding a new approach because ... you can't personally think of any alternative so the provably wrong approach must be correct?
In other news, the amount of progress into AI research depends on what fronts you judge the progress and there have been numerous steps forward but none of them have resulted in C3P0 style robots because that is not the goal.
Oh, okay
Re: (Score:2)
I don't have one. Of course, neither does anyone else. That said, beating on long disproved approaches isn't exactly going to get us anywhere. Computationalism is as dead as spontaneous generation. You don't need an alternative to find out that something doesn't work, and is never going to work. You'd have us repeat the same failure over and over rather than work toward finding a new approach because ... you can't personally think of any alternative so the provably wrong approach must be correct?
If you don't have a better plan, then it doesn't hurt to keep working on the old one. How long have people worked on human powered helicopters before they finally had some success ? I don't see you offering any fundamental reason why computationalism is dead.
Mammalian vs Human Brain (Score:2)
FTFY
A mammalian brain is not necessarily a human brain. It is the human brain we want to model.
Re: (Score:2)
This cargo-cult approach to AI is ridiculous. Decades of effort have produced absolutely no result.
Says someone who clearly hasn't kept up with recent advances in cognitive science.
Re: (Score:2)
Re: (Score:2)
Whose Brain?? (Score:2)
An interesting corollary to ponder... (Score:1)
If this approach does bare fruit (and I tend to think it will even if it requires some years yet of innovation), applying Mores law would mean we will have machines with the same number of neurons and connections as the human brain in about 30 years, and in less than 50 years such a machine will exceed the neural capacity of all humans on the planet.
Re:An interesting corollary to ponder... (Score:4, Informative)
Re: (Score:2)
applying Mores law would mean we will have machines with the same number of neurons and connections as the human brain in about 30 years
I'd love to see how you came up with this estimate.
"span the width of Manhattan"? (Score:3)
I did a double-take at that -- it just didn't sound plausible. But, sure enough, Manhattan is just a couple of kilometers wide, and a kilometer is a million millimeters. If there are millions of axons passing through that cubic millimeter of cortex, that's about how far the segments would stretch in total.
Re: (Score:1)
Also, not knowing the width of manhattan, made the whole declaration useless to me.
Re: (Score:2)
It told me it was somewhat longer than my small intestine, and somewhat shorter than my daily commute. In between, I'm not sure I'm really impressed/concerned/informed by knowing the sum length of my neurons.
Good News, Good Use of Public Funding (Score:2)
Nice to hear the government is doing some good with my taxes instead of wasting them bombing wedding on the other side of world. When I was in school, I took some AI classes and my prof had no interest in the biological neurons. Machine Learning has made some great progress, but it's still just a bag of specialized tricks. They're too brittle and don't generalize well. If we're going to ever develop a true artificial general intelligence, we're going to have to model it after our neocortex. This is a good
Re: (Score:2)
If we're going to ever develop a true artificial general intelligence, we're going to have to model it after our neocortex. This is a good start.
A good start? They've been at that for ~40 years. We're still at the 'poke it with a stick' phase. Color me skeptical.
Re: (Score:2)
50 years even. I have at home an issue of an old electronics magazine. It is from 1965, because it iintroduces the compact cassette.
In this magazine is also a schematic from an electronic neuron, built around a single transistor.
Re: (Score:2)
The sticks are getting more sophisticated. Progress is painfully slow, but still progress.
AI Research is broken (Score:2)
You aren't born knowledgable, but every AI works hard at starting from a base of knowledge. You aren't born with rules and constraints, yet every AI puts them in.
The brain is not a computer. The brain is composed of 90 Billion dumb computers that interact. Though AI wasn't powerful to follow that when Neural Nets were tri
Why not use bleach and a light microscope? (Score:2, Interesting)
There's a better way to do it, and it could potentially image the whole brain, all at once. [youtube.com] It can image whole brains of mice and other smaller mammals at the neuronal level, and we can tag each type of brain cell automatically.
Once you've got the raw data a simple AI program could map the structure logically by recognizing the tracers and plotting the connections...
Of course, this cheap and simple method may not put money in the right pockets. See what I'm thinking?
Re: (Score:2)
Imaging is only the first step. The real problem is making sense of the data you get.
Re: (Score:2)
Re: (Score:2)
There won't be much left of connections after CLARITY is done. If you want that, I think the only way you're getting it is measuring activity in a living brain. Time to crack open some rodents and wire them up.
Re: (Score:2)
Time to crack open some rodents and wire them up.
I wonder how many Manhattans of wire that would take.
how much money (Score:2)
Questionable assumption (Score:1)
Re: (Score:2)
Mind Uploading (Score:2)
Re: (Score:2)
The first uploaded organism is already done. It's a worm. It actually a composite of several worms - C. elegans has the useful feature of every individual being absolutely identical in cell layout.
Here's the worm having been Matrixed into a simulated body and environment: https://www.youtube.com/watch?... [youtube.com]
Interestingly, it swims just fine without neurons. Basic motion seems to be a function of muscle cells alone - the nervous system just determines where to go.
Laying the ground work... (Score:2)
for another AI winter.
Strong AI:
Our motto is: over-promising and under-delivering since 1951.
Our main algorithm is:
1) Remarkable step on a well defined area in AI is made (e.g. AlphaGo)
2) Issue lots of press releases
3) Claim that single isolated step is proof that all remaining thousands of steps needed are just around the corner
4) Apply for grants/create startups
5) Profit!
6) Ten years later AI winter sets in
7) A few years later, serious AI researchers who have quietly been plodding along make another remar
Re: (Score:2)
And consider this - mightn't a sufficiently complex algorithm appear to be non-algorithmic?
Re: (Score:1)