Some neuroscientists believe it may be possible, within a century or so, for our minds to continue to function after death — in a computer or some other kind of simulation. Others say it’s theoretically impossible, or impossibly far off in the future. A lot of pieces have to fall into place before we can even begin to start thinking about testing the idea. But new high-tech efforts to understand the brain are also generating methods that make those pieces seem, if not exactly imminent, then at least a bit more plausible. Here’s a look at how close, and far, we are to some requirements for this version of “mind uploading.”
Preserve the brain
The hope of mind uploading rests on the premise that much of the key information about who we are is stored in the unique pattern of connections between our neurons, the cells that carry electrical and chemical signals through living brains. You wouldn't know it from the outside, but there are more of those connections — individually called synapses, collectively known as the connectome — in a cubic centimeter of the human brain than there are stars in the Milky Way galaxy. The basic blueprint is dictated by our genes, but everything we do and experience alters it, creating a physical record of all the things that make us US — our habits, tastes, memories, and so on.
It is exceedingly tricky to transition that pattern of connections, known as the connectome, into a state where it is both safe from decay and can be verified as intact. But in recent months, two sets of scientists said they had devised separate ways to do that for the brains of smaller mammals. If either is scaled up to work for human brains — still a big if — then theoretically your brain could sit on a shelf or in a freezer for centuries while scientists work on the rest of these steps.
Both start by “fixing” the structure of a brain in place with a chemical similar to that used in funeral homes for embalming.
The first method involves soaking the brain both in osmium, a heavy metal that stains the outlines of all the neurons so they can be seen under an electron microscope, and a special-sauce solvent that allows the stain to penetrate evenly. Water is drained from the brain, which is then embedded in a hard plastic resin so that the brain’s connections can be scanned later on.
The potential problem is that all the staining and draining will obscure the molecular identity of each neuron and synapse — another critical piece of information likely necessary for a brain simulation.
The second method involves storing a brain at cryogenic temperatures (that is, really, really, really cold). Historically, the problem with that has been that it allows the formation of ice crystals that can slice through neurons and shatter synapses. A kind of antifreeze or “cryoprotectant” can help prevent ice from forming, but that in turn can lead to dehydration, shrinking the connectome so that it cannot be seen under an electron microscope, and potentially causing it to tear.
Scientists seem to have found a way around that — at least in pig brains — by fixing the structure in place and adjusting the cocktail of chemicals pumped through the brain. This method is far more likely to preserve the molecules. But how to stain and embed the brain in plastic would still need to be figured out before moving on to the next step.
Preserve the brain
The hope of mind uploading rests on the premise that much of the key information about who we are is stored in the unique pattern of connections between our neurons, the cells that carry electrical and chemical signals through living brains. You wouldn't know it from the outside, but there are more of those connections — individually called synapses, collectively known as the connectome — in a cubic centimeter of the human brain than there are stars in the Milky Way galaxy. The basic blueprint is dictated by our genes, but everything we do and experience alters it, creating a physical record of all the things that make us US — our habits, tastes, memories, and so on.
It is exceedingly tricky to transition that pattern of connections, known as the connectome, into a state where it is both safe from decay and can be verified as intact. But in recent months, two sets of scientists said they had devised separate ways to do that for the brains of smaller mammals. If either is scaled up to work for human brains — still a big if — then theoretically your brain could sit on a shelf or in a freezer for centuries while scientists work on the rest of these steps.
Both start by “fixing” the structure of a brain in place with a chemical similar to that used in funeral homes for embalming.
The first method involves soaking the brain both in osmium, a heavy metal that stains the outlines of all the neurons so they can be seen under an electron microscope, and a special-sauce solvent that allows the stain to penetrate evenly. Water is drained from the brain, which is then embedded in a hard plastic resin so that the brain’s connections can be scanned later on.
The potential problem is that all the staining and draining will obscure the molecular identity of each neuron and synapse — another critical piece of information likely necessary for a brain simulation.
The second method involves storing a brain at cryogenic temperatures (that is, really, really, really cold). Historically, the problem with that has been that it allows the formation of ice crystals that can slice through neurons and shatter synapses. A kind of antifreeze or “cryoprotectant” can help prevent ice from forming, but that in turn can lead to dehydration, shrinking the connectome so that it cannot be seen under an electron microscope, and potentially causing it to tear.
Scientists seem to have found a way around that — at least in pig brains — by fixing the structure in place and adjusting the cocktail of chemicals pumped through the brain. This method is far more likely to preserve the molecules. But how to stain and embed the brain in plastic would still need to be figured out before moving on to the next step.
For now, the only way to see all the connections in the three-dimensional maze of wiring hidden in brains is to zoom in on brain tissue with an electron microscope and scan it in extremely thin sheets, one after another, then stitch the scans back together in a computer.
To avoid missing any of the densely packed pathways that may encode a memory of, say, how your grandmother's chocolate cake smelled emerging from the oven, researchers must carve a single cubic millimeter of brain into some 30,000 slices — and a human brain contains no fewer than a million of those cubic millimeters.
The advent of new tools has automated this arduous process, which is now relatively reliable and, while still painfully slow, faster than ever before. The most ambitious such project now underway, a scan of an entire mouse connectome, is expected to take about five years using the fastest (and most expensive) electron microscope in the world. At that rate, it would take thousands of years to scan an entire human brain. But one recent paper says that it might be possible to parcel out chunks of brain to many such microscopes working in parallel without compromising the accuracy of the final map.
If that worked (another big if), and if a super-thin-slicing device could be invented (the diamond knives used for smaller brains probably wouldn’t work for human-size brains), and if someone wanted to spend some $3 billion on fancy microscopes, you could maybe scan a whole human brain in 25 years.
Trace the connections
But you still wouldn't be anywhere close to a brain simulation.
To have any hope of simulating brain activity, the mountain of scans needs to be transformed into models of the brain’s wiring. That means connecting the parts of hundreds or thousands of different layered images that correspond to a single neuron (the people who work on this call it “tracing”) and pinpointing every spot where different neurons form a synapse (about 10,000 on average).
This turns out to be even harder than it sounds, because today's computers, which can beat human chess grandmasters, can't reliably recognize on electron microscope images where one neuron ends and another begins: they need human tutoring and supervision. As a result, it took Harvard University neuroscientists thousands of hours to annotate the tiny fragment of a mouse brain showcased on the cover of the science journal Cell this summer.
Still, advances in machine learning have made coloring in connectome maps some fifty times faster over the last five years, according to one published measurement. The pace is expected to pick up, thanks in part to other applications of so-called deep learning, like teaching self-driving cars to recognize pedestrians. And after being taught to trace brain circuits devoted to object recognition, such machines may ultimately — and eerily — uncover the clues required to improve their own performance.
Interpret the map
The real challenge for aspiring mind uploaders will be figuring out how to create a fully functioning model of a human brain from a static snapshot of its connectome. To work, that model would have to include the molecular information in its neurons and synapses. Many neuroscientists think extracting that information would require another major step, others say structural details visible in the electron microscope might allow them to infer it.
But there are other challenges, like devising and testing theories about how it all works.
Imagine looking at the wiring diagram of a radio with no means to power it on, except that instead of a radio it's the most complicated machine ever invented. You can see the wires but you don't know the function of the components they connect, and unlike with electrical circuits, there are an as-yet-untold variety of them.
That's about where we stand now. In 1986, the roundworm C. elegans, which has a mere 302 neurons, was the first, and to this date only, animal to have its complete connectome mapped (the researchers who did this won a Nobel Prize). Yet the brain of the worm, whose behaviors amount to smelling food and moving toward it, has never been successfully simulated.
Human brains have about 100 billion neurons, so the enormousness of the task for a human brain is hard to overstate.
But some progress is being made — enough, anyway, so that the Obama administration signed off last year on a request by the National Institutes of Health for $4.5 billion to deliver a “comprehensive, mechanistic understanding of mental function” by 2025. Private foundations, like the Allen Institute for Brain Science and the Howard Hughes Medical Institute, have also announced major investments in basic brain research in recent years. And this summer, the blue-sky research arm of the United States intelligence agencies, Iarpa, distributed some $50 million in five-year grants to map the connectome in a cubic millimeter of mouse brain linked to learning behavior, record the corresponding neurons in live mouse brains and simulate the circuits in a computer.
To have any hope of simulating brain activity, the mountain of scans needs to be transformed into models of the brain’s wiring. That means connecting the parts of hundreds or thousands of different layered images that correspond to a single neuron (the people who work on this call it “tracing”) and pinpointing every spot where different neurons form a synapse (about 10,000 on average).
This turns out to be even harder than it sounds, because today's computers, which can beat human chess grandmasters, can't reliably recognize on electron microscope images where one neuron ends and another begins: they need human tutoring and supervision. As a result, it took Harvard University neuroscientists thousands of hours to annotate the tiny fragment of a mouse brain showcased on the cover of the science journal Cell this summer.
Still, advances in machine learning have made coloring in connectome maps some fifty times faster over the last five years, according to one published measurement. The pace is expected to pick up, thanks in part to other applications of so-called deep learning, like teaching self-driving cars to recognize pedestrians. And after being taught to trace brain circuits devoted to object recognition, such machines may ultimately — and eerily — uncover the clues required to improve their own performance.
Interpret the map
The real challenge for aspiring mind uploaders will be figuring out how to create a fully functioning model of a human brain from a static snapshot of its connectome. To work, that model would have to include the molecular information in its neurons and synapses. Many neuroscientists think extracting that information would require another major step, others say structural details visible in the electron microscope might allow them to infer it.
But there are other challenges, like devising and testing theories about how it all works.
Imagine looking at the wiring diagram of a radio with no means to power it on, except that instead of a radio it's the most complicated machine ever invented. You can see the wires but you don't know the function of the components they connect, and unlike with electrical circuits, there are an as-yet-untold variety of them.
That's about where we stand now. In 1986, the roundworm C. elegans, which has a mere 302 neurons, was the first, and to this date only, animal to have its complete connectome mapped (the researchers who did this won a Nobel Prize). Yet the brain of the worm, whose behaviors amount to smelling food and moving toward it, has never been successfully simulated.
Human brains have about 100 billion neurons, so the enormousness of the task for a human brain is hard to overstate.
But some progress is being made — enough, anyway, so that the Obama administration signed off last year on a request by the National Institutes of Health for $4.5 billion to deliver a “comprehensive, mechanistic understanding of mental function” by 2025. Private foundations, like the Allen Institute for Brain Science and the Howard Hughes Medical Institute, have also announced major investments in basic brain research in recent years. And this summer, the blue-sky research arm of the United States intelligence agencies, Iarpa, distributed some $50 million in five-year grants to map the connectome in a cubic millimeter of mouse brain linked to learning behavior, record the corresponding neurons in live mouse brains and simulate the circuits in a computer.
Correcting errors
It’s not really clear what this step would involve, but some neuroscientists say broken connections might be pieced back together. And molecular information may hold clues to aid in reconstructing imperfections in the wiring diagram. No one knows how much missing information would be acceptable in reconstructing who you are — but then, how much of you do you really need to be you?
Write a simulation
Compared with the other steps, neuroscientists seem to think this part would be pretty easy, maybe because they're not computer scientists. The European Union's billion-dollar Human Brain Project, the largest-scale simulation yet, has been criticized as being ahead of its time. That is, it is basing its work on assumptions on what it is supposed to be simulating, rather than data.
But computer simulations of a few hundred neurons here and there are widely used in neuroscience to test theories against reality. There is currently no computer powerful enough to simulate the whole human brain, but “exascale” computers, believed to match the brain's computational speed, are supposed to be introduced in a few years. How quickly a reasonably sized device could power a human brain simulation will depend on computing power, whose celebrated trend of doubling every 18 months has lately been said to be in doubt.
Connect to a robotic or virtual body
At this point in the conversation, some of you might be wondering if you’re not already a simulation. But if the idea of simulated life troubles you, consider that robots have gotten fairly sophisticated over the last 10 years — they can run, jump, see and hear, and some even play heavy metal music.
If your mind can fit on a computer chip of the future, it could, the speculation goes — and this really is entirely speculation now — be stored in the cloud of the future and used to control a robot body, or attached to one. On the other hand, some middle schoolers spend all their time in Minecraft, often playing with people miles away, and Oculus Rift, the virtual reality company purchased awhile back by Facebook for $2 billion aims to put tens of millions of people into virtual worlds. Maybe you’ll prefer the virtual to the physical, or maybe you’ll switch back and forth.
Test whether it is really ‘you’
This is obviously the most important part, but there’s really no way to know whether a faithful copy of your brain would be you. Philosophical debates continue to be waged about whether mind uploading would actually transfer your consciousness or just make a copy of you, which might be nice for your friends and family members, but you'd still be dead. There is no data on which to base a conclusion. Some people think it's obvious that your upload would be you, others think it's obvious that it wouldn't.
Neuroscientists seem as divided on this as regular people. What if your upload acts exactly like you, has all your memories, and says it is you? How would you tell? What if 50 copies are made of that copy of you, and they all say they are you too?
But maybe at this point we’re getting ahead of ourselves.
http://www.nytimes.com/
No comments:
Post a Comment