Towards a Roadmap of Whole Brain Emulation
Posted by Jeriaska on August 12th, 2007
Anders Sandberg at Transvision 2007
Dr. Anders Sandberg is a Swedish neuroscientist, science debater, futurist, transhumanist, and author. He earned a Ph.D. in computational neuroscience in 2003 from Stockholm University. He is currently researcher in the Oxford group of the EU ENHANCE Project at the Uehiro Centre for Practical Ethics and research associate at the Future of Humanity Institute (both part of Faculty of Philosophy, Oxford University).
He is cofounder of and writer for the think tank Eudoxa. Between 1996 and 2000 he was Chairman of the Swedish Transhumanist Association. He has also been scientific producer for the neuroscience exhibition “Se Hjärnan!” (“Behold the Brain!”), organized by Swedish Traveling Exhibitions, the Swedish Research Council and the Knowledge Foundation that is touring Sweden 2005-2006. His presentation at Transvision 2007 was called “A Roadmap Toward Whole Brain Emulation” and subtitled, “Why Philosophers Actually Care About Microscope Jitter.”
The following transcript of Anders Sandberg’s Transvision 2007 presentation “Towards a Roadmap of Whole Brain Emulation” has not been approved by the author.
Towards a Roadmap of Whole Brian Emulation
It seems that the first three speakers here are philosophers. Or, rather, I am not a philosopher, I just play one at Oxford. I’m going to show what kind of weird course of philosophy I am doing, because traditionally we have a view, what is philosophy? It is wise men sitting in an office somewhere and thinking deep thoughts. But it is usually quite disconnected from reality. Of course you have practical philosophy, which is supposed to be connected to reality a bit more. But one of the interesting problems when we get practical philosophy as it relates to transhumanism (which involves issues of bioethics and biopolitics, should we be allowed to enhance ourselves) is that the questions that come up are often quite practical.
For example, if I were to say that we should increase IQ as much as possible in the population, some people say that’s a bad idea, it is going to increase inequalities. That is actually a practical question. Similarly, one of the favorite questions for a lot of us is, of course, how can we get rid of our bodies? Whether we think that is a good or bad idea, it has a lot of philosophical importance. Unfortunately, philosophy cannot itself answer that. We have to get into a bit more of the practical details. So, I am going to make a bit of an advertisement here for one of the workshops we did at Oxford a few months ago about whole brain emulation. We might subtitle this talk, “Why Philosophers Actually Care About Microscope Jitter.”
Well, what is whole brain emulation? Many of you know of this concept often called ‘uploading.’ Sometimes ‘downloading’ if you are pessimistic about it. The idea of transferring the mind, or consciousness, or what have you, from a biological brain into software. There are a lot of interesting reasons to think about this. From a philosophy standpoint, this is an old classic, of course. A lot of quite famous philosophers have debated the possibility, not because we are actually interested in computers or achieving immortality, but because it creates interesting questions about what is the nature of mind. What is consciousness, really? You can do a lot of wonderful thought experiments. But I think the best way to find out these things about the mind and reality is to actually try them out. As you can hear, I’m very much of a British empiricist here.
Similarly, there are a lot of interesting ethical issues surrounding enhancement. And, also, uploading might have a lot of importance there. From a scientific perspective, I am after all a neuroscientist, being able to replicate the brain is a useful research area. We can learn a lot about the brain. We can reverse engineer parts that are doing something useful. And, of course, for a focus of a lot of computational neuroscience, it can actually be very stimulating. So there are good scientific reasons to look at what is the scientific possibility of this. Is it just handwaving and errant nonsense, or is it something we should perhaps consider a bit?
Robin Hanson has written one of the most interesting and frightening essays I have ever read about the economics of whole brain emulation, where he points out that if you can copy an intelligent mind, that is going to have a tremendous economic impact. If something very powerful is likely to happen in the future, we ought to think about it and find out its probabilities. Whether it is an opportunity or a threat, we want to know ahead. Even if it is decades ahead, we might need decades to prepare ourselves. And of course transhumanism is interested in this too, whether to get rid of boring flesh, or for more practical reasons, because it might be useful to have back-up copies.
But the big problem here is, how do we actually talk about this? We are talking about a fundamentally futuristic technology that we do not have today. We do not even have anything similar. And that is, of course, another reason to try and figure it out. And that is why people in the philosophy department are interested. We are trying to figure out better ways of thinking about the future. Now, I am going to laregly ignore the philosophical issues in this talk. So I am not going to try to convince you that I think a copy of my brain in a computer, if it is good enough, is going to be me in another sense. Or whether computers can be conscious, and so on. We can have another conference on those questions. I am going to try to figure out, can this be done?
The basic issue with whole brain emulation is that we do not need to understand really the brain in order to perhaps build a record of it. It is a but like a machine. If you can get a blueprint and the components, you can actually put them together into the functioning machine, even if you have not the faintest clue about what it is doing. Those little black chips on the screen, we do not need to know what logic gates they involve. If we just put them together in the right way, it is hopefully going to work. And, the assumption here, and this is one of the main assumptions, is that if you put the right parts together, you will get a functional brain. You can of course think about a physical replica of the brain, but we will talk about a software replica. Te assumption is that if you simulate the neurons together, you get an activity in the neural network that will correspond with the activity in a real brain.
So, the nice thing is that we do not really need to understand what, say, memory is, or how we make an intelligent decision. We just need to put the parts together. And of course if you go to a sufficiently low level, the parts are going to be pretty simple. But, in principle, we could make a simulation based on atoms. We would need a method of scanning all the atoms in the brain, and then we would need to understand how atoms interact, which is fortunately pretty simple. The downside is, of course, is that it is going to require a tremendous amount of computing power to run it. We could try to do something less complex or use a high level description. The really interesting question is that we don’t know yet which level of description would be suitable and what would actually work. So that is something we are interested in finding out.
So, overall, if we look at the problem of level of detail, we cannot imagine making a complete computational model of the entire brain. That would essentially be an artificial intelligence. That would probably be pretty hard to do. We haven’t even got the faintest idea how to make an artificial intelligence in general, let alone one that mimics the entire brain. On the other hand, if we go down one level, we can look at the operations of each brain area. If we know the connectivity between them, then we can describe the function. The problem is that a single brain area does a tremendous amount of stuff. And one of the embarrassing truths in neuroscience is that the visual cortex is not only doing visual stuff. There are actually taste responsive cells in the visual cortex. It is terribly embarrassing, but true.
On the other hand, we could go down a couple levels. We could go down to level 4, where we have a simplified neural network, similar to the software for predicting stock markets. We have a neuron, but it is not a real neuron. It is highly simplified and without connectivity between the neurons, and hopefully that can simulate the brain. Now we probably have a better chance of simulating a brain, but it still does not capture completely what is going on in our brains. When we think about our mood swings, our tiredness when we wake, the peculiarities of the onset of happiness or sadness, a lot of that may actually be due to particular chemical properties or the electrochemistry of the neurons.
We know that there is a lot of computation going on down at the dendrites, the branches of the neurons. So we might need to go down to level 5, simulate smaller parts of the neurons, the individual synapses. Or we might need to go down to level 6, what chemicals are around. It is known that there are from a hundred to a thousand chemicals involved in each synapse. A lot of them are probably involved in learning and memory, and if we don’t model them in the right way, it might turn out that we get the wrong kind of learning. But of course it could be even worse. It could be that we need to go down to level 7, dealing with the different proteins, or the state of the proteins. Some proteins like to cluster together, we might need to model that. Or we might have to go down to horrible level 11, and most neuroscientists don’t believe it, that there might actually might be some quantum effects going on in the brain. In that case it would probably be impossible to do a whole brain emulation. Because it is probably pretty tricky to get a quantum state out of anything.
So, what we did was we invited a group of researchers who are doing work on scanning, simulating neurons, building computers, and we started to pick their minds in order to get an estimate of how likely is this to happen. So here is a rough outline of how a brain emulation would work. So, we take a brain, and then we scan it somehow. We really need to figure out what kind of resolution we need, how much detail, and what method we would use to scan it. Then we would construct a brain model from this data, and somehow we need to infer not just the structure but also the relevant low level information so that we can get the right component parts for a simulation. And then we set up a simulation in a sufficiently advanced computer. We might want to add a simulated environment to this. Because, if you think about a human brain (or for that matter a lab rat, which is much more likely to be the first thing that is uploaded, unless a nematode beats a lab rat) sends signals to the senses, so it needs simulated senses. At least so we could test whether it is working properly. A first round without any senses might produce some activity, but it would not really have any meaning. If you have a simulated body, or for that matter a real body in the physical world, you can see if the behavior affects the work.
There is another interesting point. This really shows the benefit of bringing real scientists to talk with philosophers, because we discovered at that meeting that we had completely different views about validation. How do we actually test that we are doing something properly? In terms of the particular brain we tried to emulate, have we actually succeeded? That is a really tricky question. In doing these methods we need to develop tools to figure out whether we are making things better. This is a very practically and philosophically interesting question. That got a lot of the philosophers thinking. So this really shows the benefit of getting engineers in touch with philosophers. Because quite often it is the philosophers who get their ideas from the engineers, not vice versa.
So, how can we do a neural simulation? Hodgkin and Huxley did some heroic work in the ’40s and ’50s in understanding how the neurons work. And they actually did the first neuron simulation by using a hand-cranked electromechanic calculator. It took them several hours to simulate a millisecond of neural activity. They had to solve differential equations by hand. They eventually managed to plot a single spike of a neuron firing, and they got the Nobel Prize for that. I think that was well worth it, both in terms of effort and also as a start on an entire field, computational neuroscience.
So basically the idea is that if you have a neuron you can divide it into smaller parts. This is the standard method. And these compartments can be regarded as rather complex electrical circuits. And then we can simulate these electrical circuits using standard numerical software. Doing this in the right way, of course, that is what keeps computational neuroscientists getting salaries and organizing conferences all the time, because there is an awful lot of interest in the complexity there. It turns out that we have pretty good neuron simulations, good enough to fool experts. They cannot tell if it is a simulated neuron or a recording from a real one. We have some 1:1 simulations of small systems. The classic one is the lobster somatogastic ganglion, a small group of cells regulating how it is swallowing. And a much larger one is a swimming network in a laboratory with a fish, in which we have about 100,000 neurons. That is more or less 1:1. It is the same number of neurons and they are likely connected in the same way as in the real animal. And it does produce a variety of activity. Experts have difficulty differentiating the simulated activity from the real one.
In the animal it seems that some other requirement would be needed, of a time resolution less than one millisecond, probably on the order of a tenth of a millisecond. We need compartments at least less than ten micrometers, at least for the really messy branching neurons. But that’s pretty doable. Right now I can say that my former officemate holds the world record for the largest realistic neural network simulation. So they did a 22 million neuron simulation, a relatively small set, just six compartments and 11 billion synapses. They ran this on one of the big IBM Blue Gene computers and it’s actually pretty impressive. It was just running about 5000 times slower than a real brain would. And 22 milion, that’s about 1/5000 of a small brain area. So we’re getting there. We’re getting there. The problem is that the network was randomly connected corresponding to an abstract idea of how the brain is organized. So it didn’t do any real thinking. But we are starting to get an estimate from this of how much computer power is required to et realistic activity.
Then of course we can complicate things. So we were discussing at our workshop a lot of the gravel that could get into our machinery. There are some issues, I don’t have the time to go through them, that seem to be just annoyances. We might have to scan the spinal chord to make a really good brain emulation. There are some things that are going to be messy to work with, like the spreading out in the volume of certain chemicals. There are some things we don’t even know yet whether they are important, like electromagnetic interactions when neurons are firing next to each other. They seem to be part of a certain disease and pain, but we do not know whether they are useful in the normal, healthy neurosystem. And there are always those physicists who think that there should be some quantum going on.
So there are interesting research issues here and we have been trying to take them into account. The next part is of course doing scanning. In general we don’t know yet what resolution we will need. One of the purposes of this workshop was to see if we could tease a consensus out of the modelers of what we need. And of course we can try to scan brains in certain ways. Most people probably have a thought that there should be such a thing as putting a head in a brain scanner and it does something magical like an MRI, and then you get an emulation. The problem is that it is pretty unlikely that we could do a non-destructive upload that way. In general, we have pretty severe limitations on how high resolutions we can get over long distances. And since a non-destructive method has to be outside the head, it does not do very well. Gradual replacement is very popular, but it turns out that that is probably very technically complex. So, from a practical standpoint, the first kinds of brain emulation are going to be based on a destructive scan.
So you would fix the brain, either freezing it or putting it into some form of plastic, and yes this would definitely kill it in a traditional sense. It will not be that easy to get volunteers for this. And then we need to find a way to get a high resolution scan. This is actually where magnetic resolution scanning can do some good. There is a limit on using MRI for a living brain because it’s not wobbling around because of the blood flow, but if you freeze it you can scan it at a ridiculously high resolution. And then you have various interesting methods of slicing it. So up there you can probably not see all the details, it doesn’t matter. Or various electron microscope methods of slicing and scanning to get the three-dimensional sense of tissue.
Now here are some illustrations from Bruce McCormick’s very interesting work from knife edge scanning, where he has a knife made out of diamond that slices the surface of a sample, and then they image it as slices. And it’s moving back and forth. And in the course of a few hours it can scan an entire mouse brain. Unfortunately, it is not scanning all the neurons, it is just scanning labeled neurons that are colored, but it’s a start. And it can be developed further. Kenneth Hayworth is doing something really cool, and I really recommend visiting his website Extreme Neuroscience, where he shows that you can gradually slice up a tissue sample in plastic to get a long strip, and then put that strip on a metal plate, and store it. And then you can use automated techniques for rolling that tap back and forwards in front of an electron microscope. So, essentially, you end up with a library of small slices that you can store likely for a pretty long time.
So it seems that optical microscopy can scan with a certain resolution limited by the wavelength of light. That might be enough, but we don’t know yet. Electron microscopy is very likely to scan whatever resolution we nee, but there is interesting trickiness here about what information we need to get from it, so we had a bit of interesting consensus discussion about that. After we have done that, we need to do post-process image processing and figure out how to process particular neurons. Again, some research is done here so we can calibrate a bit, and it seems to be moving along nicely, at least to get to the morphology: what neurons are there and how are they connected.
The big problem is that we don’t know how to get chemistry out of it. And here we discovered our first really big research question, and whoever manages to answer it will probably get a prize or something. How much can we infer about the function of a neuron just by looking at electron microscopy pictures of it? Or is there an easy way to figure out whether a neuron is a fast-spiking or a slow-spiking neuron from the picture. Nobody knows presentl, and that is probably one of the important parts to making a roadmap to brain emulation.
Another thing is of course computer power. We all want power, whether it’s for computer games or for emulation. I’ve been having a lot of fun plotting Moore’s Law, and in general you can make these nice curves. I’m certain Ray is going to show us a lot more of these curves. So, it’s all long diagrams with straight lines corresponding to exponential increase. Whether an exponential increase would continue indefinitely is another matter.
In general, the performance amount of computer power per dollar is increasing a lot. This is another data set showing roughly the same effect. Another nice thing, of course, is the amount of computer memory you can get for a dollar. It’s also increasing exponentially according to a much nicer curve. And then of course the speed of memory is getting better. So the latency, the amount of time it takes to retrieve something is getting faster, though at a relatively slow rate. Similarly, hard drive space is increasing at an amazing speed, but the spreed of hard drives is not increasing quite as much.
If we look at the time it takes to get ten times better in terms of computer power, it’s about 5.1 to 5.6 years. And we’re getting a lot more memory each year. If the computer that is running the upload is similar to ours, and not an entirely novel device, it will probably have lots of processors with a lot of local memory. It turns out some other research from a colleague of mine shows that we don’t need to worry too much about bandwidth, at least for some particular kinds of models we are using. The green line corresponds to people’s estimates of the computer power required to emulate the human mind. And one of the problems with these ones is that most of it is complete handwaving. But we can plot the trends. The little red line corresponds to what I can get for $1 million. And the upper one corresponds to what I can get for $1 billion. At that point is the earliest possible time you could do a computer brain emulation. And it seems that we could almost do it today if we had an American defense budget or something like that.
So today, we do not have the computer power, we don’t have a scan, and we don’t have the neuroscience. But in 2060, it seems pretty likely that we are going to definitely have computer power enough. So it is interesting to try to benchmark how much better microscopy is getting. And that turns out to be much trickier. Then there is the simulated environment, and I don’t really have time to get into it, but it is pretty interesting. A lot of work is being done there with computer games, virtual reality, and actually medical imaging. It turns out that making a body simulation and making a world simulation just takes 100 teraFLOPS. Quite a lot today, but it’s going to be trivial if we can run a brain emulation.
So, in general, we can try and make a more complex estimate than just make assumptions. You remember the levels of detail I was showing you earlier? If it turns out that we really need to simulate things down to level 9, in that case we can look at how many things of that kind are in the brain, because we need to scan them, store them, and simulate them. And then we can use previous estimates to actually get the first date. So, if we could do everything on level 3, we could use pretty simplified neurons, in that case we would need about a billion of those very simplified neurons, and you would need about 10,000 bits per one of those. So you would need 200 terabytes of storage. And in order to get that for $1 million, you can do that in 2010.
That’s not bad. The problem is that simulating is probably going to take 10^15 FLOPS, and you are not going to get that for $1 million until 2033. On the other hand, if it turns out that you need more detailed simulations, like level 5, then you are going to need 10^14 compartments of neurons, and you can still get memory storage for it by 2020. But it is going to take until 2046 when you can actually do the simulation.
So the interesting thing here is which road we are on, that is going to be determined by nature. That’s not up to us. We have to make a kind of gamble on it. What we can determine is of course where we can do research, and we can try to calibrate it. We can actually update these estimates as we get better at it and learn more. So this is a diagram that you can imagine would dynamically give ongoing estimates.
To sum up things here, what did we find? Well, the rough consensus seems to be that 5 x 5 x 50 nanometer resolution is probably what we need to do it. And actually most of the people in the room decided that level 5 or 6 was a conservative estimate of where we think we need to do simulations. We probably need to also simulate some weird electrophysiology of neurons, and we might need to simulate a lot of chemistry. That’s not so bad. We’re already doing a lot of that, we just need to get better at it.
We didn’t have anybody saying that you need to go down to level 11 and do quantum stuff. And we didn’t have anyone say that it was trivial, you can just replace any brain with a very short basic program. It turns out that the big question that we need to figure out, and it’s a good neuroscience question, finding out how much we can learn from the shape of neurons from what we see through an electron microscope. We also have to answer the interesting question of finding, because that is going to drive a lot of this.
Right now, the problem is that the smartest thing to scan today would probably be a snail’s brain. Not that many neurons. Animal rights activists don’t get too upset about snails. A good target, but not very sexy, because we would be talking about the grumpy brown British ones. Another target that we might want to check out is the retina. From a funding perspective, it might be smarter to try to scan a cortical column. Actually Kenneth Hayworth has an interesting idea about doing that kind of scan online, where researchers can log on to this library of scans and zoom in on any part of a small piece of tissue, and actually get it at extremely high resolution. That might actually get funding if we are lucky.
We computational neuroscientists might be a bit too slow. The electrophysiology also is not glamorous, unfortunately. And so far we haven’t gotten that much funding for large scale scanning. The nice thing, though, is that we are getting some impetus for the industrialization of neuroscience. The big thing where you have robotic labs doing genetic proteomic work, you could probably apply that to roboticized neuroscience. There is no money in it yet, but it is becoming cheaper very rapidly. And that means we are going to be able to do neuromics rather soon.
But quite generally, why should we really care about these questions? Isn’t this just handwaving for futurists? Well, suppose we manage to do the scanning and simulation rather soon. And then we can run a glorious C. elegans nematode simulation. So you’re going to get a little simulated worm crawling around your screen. That’s not going to be very dramatic, but as computers get better, you’re going to get a small simulated snail, and then a simulated mouse, and then people are starting to pay attention. Next, you get a cute simulated cat prancing around, people are going to realize that if Moore’s Law continues we are going to get to a human in 20 years, or something like that. In that case you are going to get adaptation in society.
But suppose we get the opposite situation. First the computers get better, but we neuroscientists don’t do our jobs at all. We can’t figure out how to do things. Gradually we get good scanners, but how to get a simulation right, it eludes us. But then suddenly someone gets a good idea, so then we can do the snail, and the mouse, and the human relatively soon after each other. Then you are going to get a Dolly-like panic. And that matters. We actually have a lot of leverage from it now, and I think we should make use of it.
So, I think the benefit of getting together engineers and philosophers in the room, and hopefully also the politicians, economists and social scientists, is that we can start thinking about what are our leverage points? And how can we set up gradually improving estimates of where we are going? Because we need these metrics to figure out what’s happening.
Well, I would like to thank the people who organized our brain emulation workshop, and especially the participants who actually provided the brains. And I would like to thank you for so patiently listening to a lot of neuroscience early in the morning. Thank you.
Related articles: “The Coming Merger of Human and Machine.” Ray Kurzweil outlines the foreseeable implications of accelerating technological progress.
Trackback: Sentient Developments, “You Either Get It Or You Don’t”