Perhaps because AI itself is to simulate the emergence of human intelligence, in almost every cycle of the development of AI, when we are limited by computational power or application environment, AI technology can not produce a breakthrough in efficiency, we will turn to the study of human brain, trying to simulate the way the brain works with computers.
Nowadays, although deep neural networks are used more and more widely, we can also find more and more particularities of human brain operation.
For example, in identifying animals, deep neural networks need to invest a large number of giraffe images in the black box to enable AI to identify the "giraffe Ben deer". But for human children, giraffe skeletons can be recognized by seeing a picture of a giraffe once.
This mysterious cognitive process deserves constant exploration and deduction.
Reverse Engineering of Brain Tissues of 1 Cubic Millimeter
In recent years, a project called Machine Intelligence from Cortical Networks (Microns) has provided a brand new idea for the industry: reverse engineering of the gray matter cortex, deciphering its operation strategy, and converting it into an algorithm that can be used by the machine.
This project comes from the "BRAIN Initiative" put forward by the Obama Administration in 2013. With the support of US$100 million, scientists are encouraged to study the operation of the human brain from the perspectives of cognitive science, neuroscience and fusion science.
This initiative is seen as the second Human Genome Project, which involved governments and academic institutions from many countries and took 13 years to sequence the human genome. Many people have questioned the significance of the Human Genome Project, but now it is playing an important role in genetic research.
Microns is currently the most completed project in the initiative, funded by the US Senior Intelligence Research Program. The specific research method is to draw a 1 cubic millimeter neuron structure of mouse brain tissue and study the mode of circuit connectivity between neurons, so as to reverse deduce how the animal brain responds to external stimuli.
A cubic millimeter of mouse brain tissue, compared with humans, is only one millionth the size of the human brain. Even so, it still means 50,000 interconnected neurons and 500 million synapses.
We know that reverse engineering means re-deducing the birth process of a product after the final shape of the product is known. So how to carry out "reverse engineering" in the face of such a huge problem?
From microscopy to DNA, what are the ways to record neuronal movements?
The American Advanced Intelligence Research Program (AIRP) has chosen to work with three research teams in three ways to study the brain tissue of a cubic millimeter mouse.
Harvard University chose electron microscopy. By injecting fluorescent protein into rats and training them, the brain activity is stimulated by playing video for rats. When neurons are active, calcium ions in fluorescent protein will be incorporated into the cells to make them shine. At this time, the activity of neurons was recorded by laser microscopy. On the other hand, a cubic centimeter of brain tissue is cut into thin slices and imaged under a high-resolution microscope. By comparing and mapping the status of active neurons with that of intact inactive neurons, the "mode of thinking" of experimental rats can be excavated.
Experts from Harvard Medical School chose another method, labeling neurons with a special DNA bar code, which identifies the motor of neurons. As for brain slices, the information can be classified by gene sequencing machine to reproduce the motor state of neurons.
The team from the American Association for the Advancement of Science simply chose a data-driven approach to build the research base by recording the connections of brain neurons in a comprehensive way.
In the plan, the three teams will monitor the movement of tens of thousands of neurons in the brain together, and compute and stitch the cross sections of brain tissue slices to connect the active paths of neurons to form a three-dimensional map of brain movement. Based on this, we try to simulate the pattern of neuron movement.
Paradox in Reverse Engineering: Realizing Fewer Samples with Big Data?
In this way, the reverse engineering for the brain is not different from the previous brain simulation engineering and so on, besides being more specific in research methods and focusing more on details in division of labor.
But it's worth noting that the biggest difference between Microns and other brain simulations is that they have clear goals.
When the BRAIN Initiative was launched in 13 years, the goal was to study Alzheimer's disease, autism and other diseases by studying the brain. But when it actually started, Microns, the most popular project, targeted AI applications, each of the three teams with at least one algorithm scientist, in order to translate neurological results into applied computer science.
In this way, there is a feeling between brain research and artificial intelligence that "nothing happens, nothing happens in summer and spring". Usually, brain science research always aims at psychology, neurology and medicine, but once artificial intelligence is hot enough, brain-like computing, cognitive computing related to brain science and so on, it immediately becomes the bright light of artificial intelligence.
In the case of Microns, there is a serious paradox.
First of all, the goal of Microns is to achieve more efficient learning with fewer samples or even without samples by simulating and deducing the brain, so that the neural network can no longer rely on a large amount of data to build models.
As far as the current method of brain reverse engineering is concerned, it is not the algorithm that Microns first output, but the huge amount of neuron motor data, which can produce 1 to 2 PB data per cubic millimeter of brain tissue.
So in order to deal with these data, Microns took the lead in developing a neural network model that can carry large amounts of data. It took a lot of time to train, and maybe it also needed to make use of overcalculation.
The whole process runs counter to the original intention of the Microns project.
Whose bricks and tiles is Microns?
In fact, researchers themselves are pessimistic about the future of Microns. Although the sponsorship cost of Microns is now as high as hundreds of millions of dollars, David Cox, a neuroscientist at Harvard University who participated in the study, said that human brain research is too complex a proposition, and they are bound to produce results, but these results are difficult to meet people's expectations.
At present, Microns has been able to classify some of the neurons of experimental mice. For example, when stimulated, identify which neurons are interconnected and which are relatively independent. However, such a result is not only far from the application, but also difficult to be theoretical and systematic.
Finally, the results of the Microns project are likely to accumulate a large amount of data on neuronal movement, waiting for more forces to work together after opening up to society. This process is like building a house. Perhaps what we are witnessing today is just the preparation of bricks and tiles. Maybe it's artificial intelligence or brain science that makes use of these bricks and tiles.
The relationship between brain science and artificial intelligence, sometimes like the heroes and heroines in dog's bloody love stories, seems to be a natural pair, but after twists and turns, they are always unable to be together. But in the process of chasing each other, they all got better growth.