(Original title: Inside The Moonshot Effort to Finally Figure Out The Brain) NetEase Technology News, October 18, according to foreign media reports, at present, artificial intelligence is only a loose imitation of the brain.
so what if you really imitate a brain? What you need to do so far is completely impossible: map all the neurons and nerve fibers in the brain.
is the problem that artificial intelligence is facing right now," said David Cox, a professor at the University of New China.
" yes, artificial intelligence has developed very well, from near-perfect facial recognition to self-driving cars to defeating the world go champions.
, some AI applications don't even need to be programmed: their architecture allows for continuous self-learning through experience.
but Cox, a Harvard neuroscient scientist, says artificial intelligence still has some lame problems.
To develop a puppy detector, you need to enter thousands of dogs into the program, as well as thousands of animals that are not dogs," he said.
and the e-knowledge of ai's experience extracted from all the input data is extremely fragile.
even some people will not notice the image noise will bring a lot of trouble to the computer, such as let the computer trash can as a dog.
the effects of facial recognition on smartphones, such as security, are not very good. to overcome this limitation,
last year Cox teamed up with dozens of neuroscientists and machine learning experts to launch a project called Cortological Neural Network Machine Intelligence (MICrONS), a $100 million project to rebuild the brain through reverse engineering.
Vogelstein, an official with the Advanced Intelligence Research Program, a U.S. research organization, is the creator and founder of the MICRONS project, which he calls the Apollo program in the neuroscience community.
(he is now a partner at a venture capital firm in Baltimore).
the MICRONS project are trying to map out the structure of a small piece of the rodent's cerebral cortation, restoring the function and structure of every detail in it.
fact, every cubic millimeter, with just one gravel-sized map of the brain cortation, is a moon landing project.
the details of the map that people want to depict are hundreds of millions of times its size.
it contains about 100,000 neurons, in addition to about 1 billion synapses, by which nerve impulses leap from one neuron to the next.
project's ambitions have left other neuroscient scientists in awe.
I think they've done a very heroic thing," said Eve Marder, who has contributed to it throughout her career.
one of the most exciting things about neuroscience, " says Konrad Kording, who studies brain computational models at the University of Pennsylvania in New York.
the ultimate reward of "research", as Vogelstein points out, is to form "the computational basis for the next generation of artificial intelligence" by mining the neural secrets behind the project's data principles.
that current neural networks work on structures that were proposed decades ago, based only on a fairly simple conceptual brain nerve, " says Dr. Vogelstein.
In essence, artificial intelligence systems spread information to thousands of closely connected "nodes", each of which is similar to neurons in the brain, and the entire system improves system performance by constantly adjusting the strength of the connections.
but in most existing computer neural networks, signals are always passed from one set of nodes to the next cascade.
contrast, the real brain is full of feedback: each bundle of nerve fibers transmits information from one region to the next, while the same or more fibers return the signal.
but why does the brain work like this? Are these feedback fibers the secret to the brain's ability to learn powerful functions such as one-off learning? Is there any other principle? Sebastian Seung, a neuroscientist at Princeton University who was a key figure in the mapping project, said MICRONS should at least provide some answers to how the brain works.
in fact: "Without such a project, I don't think we can answer these questions." The
Focus on Details MICRONS project consists of three teams, one led by Cox, one from Rice University and Baylor School of Medicine, and one from Carnegie Mellon University, each of which is doing the same research: rebuilding all cells in the brain of a cubic millimeter mouse, while also reconstructing a map of connections between each cell and using data detailing how to stimulate neuron activity and affect other neurons.
first step in the project was to study the brains of mice to find out what neurons actually worked in the cubic millimeter brain tissue.
animals are given specific visual stimuli, such as straight lines in one direction, which neurons suddenly become active and which adjacent neurons react at the same time? In the last decade, it has been difficult to capture this data.
, "There's never been a similar tool, " admits Vogelstein.
" although researchers can implant ultra-fine wires into the brain and record the activity of individual neurons.
but because neurons are tightly aligned, researchers can't record the activity of up to dozens of neurons at once.
researchers were also able to map the overall nerve activity of the brain in humans and other animals using nuclear magnetic imaging, but were unable to monitor individual neurons in this way: the spatial resolution of MRI was as high as one millimeter.
but the development of related technologies has broken the deadlock by allowing neurons to glow when active.
to do this, scientists typically inject fluorescent proteins into neurons, which glow whenever neuron cells are active, thanks to the influx of calcium ions.
, scientists inject proteins into the brains of rodents through benign viruses, or integrate genes from fluorescent proteins into the genomes of neuron cells through genetic coding.
then, scientists triggered fluorescence in a number of ways, the most effective of which was to send infrared light through an opening in the skull of an experimental rat into the brain.
of infrared light allows photons to penetrate relatively opaque nerve tissue and be completely absorbed by fluorescent proteins.
these proteins absorb energy from two infrared photons and release one of their visible photons.
when the experimental rats saw or performed the action, the photons could be observed under a normal microscope. Andreas Tolias, head of the
Baylor team, said it was a "revolutionary technical approach" because "you can record every neuron, even those adjacent to each other."
the Cox team had mapped the neural activity of laboratory rats, the experimental animals were killed and the heavy metal molyps were injected into their brains.
then a team led by Harvard biologist Jeff Lichtman will cut the brain into thin slices to determine how neurons are organized and connected.
process will begin in a basement lab, where a desktop machine works like a sausage slicer.
small metal plates rise and fall, methodically cutting off the tip of what looks like amber crayons and pasting the slices to a conveyor belt made of plastic tape.
difference is that this amber crayon is actually a hard resin tube that wraps and supports fragile brain tissue, and the small metal plate is fitted with a sharp diamond blade, resulting in a thin slice cut only 30 nanometers thick.
, in another lab, the researchers installed tape containing several brain slices on silicon wafers and placed them in large industrial refrigerators.
the device is actually an electron microscope: it uses 61 electron beams to scan 61 brain tissue simultaneously at a resolution of 4 nanometers.
it takes about 26 hours for each wafer scan to complete.
display next to the microscope shows the resulting images, and the details of the brain tissue restored are amazing - you can see cell membranes, mitochondria, and nerve-wiring vesicles that gather on synapses.
is like focusing on a parting image: the more you zoom in, the more complex you can observe.
but the slice is not the purpose of the experiment, and the scanned image of the microscope is not the end result.
"We're making a movie that's going to stretch every slice," says Mr Leachman.
", the sliced information was forwarded to a team led by Harvard computer scientist Hanspeter Pfister.
"Our role is to extract as much information as possible from the image," says Pfister.
means using two-bit slice images to reconstruct all three-dimensional neurons in brain tissue -- including all of them, synapses and other features.
Pfizer says that while humans can do this with paper and pen, the process is very slow.
he and his team trained neural networks to track the depiction of real neurons.
"It works a lot better," he said.
regardless of the size of each neuron, it contains a variety of synapses, a large number of which are called synapses, and each neuron has a long, thin fiber called an axon.
axons are used to transmit nerve impulses over long distances, and they can even travel entirely through the brain to the spinal cord.
but using the MICRONS project to map a cubic millimeter of brain tissue, the researchers were able to track most axons from beginning to end to observe a complete neural circuit.
Think We're going to find a lot of mysteries," Pfister said.
the "expected force MICRONS team began to want to start answering the question: What is the brain's algorithm?" How do all these neural circuits work? Especially what role did that feedback play? There is currently no feedback in many AI applications.
signals in most neural networks are transmitted from one node to the next, but generally not in reverse.
(don't mix with "reverse propagation," which is one way to train neural networks.)
) Of course, this is not a fixed number: the circulating neural network does have an inverse connection, which helps the node process input over time.
, however, the feedback scale of circulating neural networks is far from the level of processing in the brain. Tai Sing Lee of Carnegie Mellon,
, points out that by delving into parts of the visual cortal layer in the brain, "only 5 to 10 percent of synapses are receiving input from the eye," with the rest listening for feedback from the next level.
two theories about feedback, cox said, "one of which is that the brain is constantly trying to predict input."
" can be said that while the sensory corty is dealing with the current scenario, other brain tissue is trying to predict the next scenario and transmit the best guesses through the feedback network.
this is the only way the brain responds to rapid environmental change.
"the neurons are really slow to process," Cox said.
" retina-sensing light can take 170 to 200 milliseconds to be transmitted to perceptive neurons.
time was enough for Serena Williams' tennis to fly nine metres.
"so anyone who wants to receive the ball must wave the racket on a predictable basis."
if you're constantly trying to predict the future, Cox says, then when the real future comes, you can adjust it to make the next prediction better.
is consistent with the second main theory about feedback: that the brain's feedback connections can be used to guide learning.
, computer simulations suggest that constant revision of improvements creates better models of the real world.
, for example, Cox said, "When a person turns around, you must know what a face will look like."
," he stressed, which could be the key to solving the "one learning" problem.
, "When my daughter first saw the dog," Cox said, "she didn't have to know how the shadows were or how the light was reflected."
" she already has a lot of experience with similar things.
" so when she sees something like "that's a dog," she adds that information to her knowledge base.
"If these ideas about brain feedback are correct, they can be fully demonstrated by a detailed map of brain structure and function drawn up by the MICRONS project."
MICRONS can prove how neural circuits can be predicted and learned.
, artificial intelligence applications can mimic this process.
, however, we still can't answer all the questions about the brain.
understanding neural circuits doesn't solve everything.
communication between cells does not depend on synapses, some of which is transmitted through hormones and free neurotransmitters between neurons.
there is also the scale of the study.
project like MICRONS is a leap forward for neuroscience, it is only a small piece of the brain's corty that has been studied to solve computational-related problems.
compared to the entire brain, the cortical layer is only the brain's thin outer layer, and key command and control functions are hidden in deep brain structures such as the alps and substrate nerves.
good news is that MICRONS is already mapping the scale for the future.