In a laboratory near San Jose, the United States, IBM built an electronic rodent brain with 48 TrueNorth test chips. Each chip can simulate a basic component of the brain.

Under the leadership of the project leader, Dharmendra Modha, we had a close contact with the whole project. It is understood that its volume is like a medicine box in a bathroom. It is covered with translucent plastic plates, which can clearly see the chips, circuit boards and colorful indicators inside. It looks like an object from a 1970s science fiction movie, but Modha said, "What you see is a small rodent."

He means the brain of a small rodent. At least this chip can support that brain. These chips serve as the basic components of the brain - neurons. Modha said that this system can simulate 48 million nerve cells, which is basically equal to the number of nerve cells in the brain of small rodents.

At IBM, Modha is responsible for the Cognitive Computing Group, which invented the "Neurochip". When he and his team first published the invention, they used it to support scholars and government researchers in IBM's research and development laboratory in Silicon Valley, and carried out a three week trial run. After connecting their computers to the digital mouse brain, these researchers explored its structure and began to write programs for the TrueNorth chip.

Last month, some researchers have seen this guy in Colorado, so their programs can make it recognize photos and speech, and understand some natural languages. The chip is responsible for running the "deep learning" algorithm, which currently dominates the Internet's AI services. It provides Facebook with facial recognition services and Microsoft's Skype with real-time language translation. However, IBM still has a leading edge in this area, because its research can reduce the demand for space and power supply. In the future, we may put this artificial intelligence into mobile phones and other small devices, such as hearing aids and watches.

"What can we get from the synaptic structure? We only need very low power consumption to classify pictures, and we can also constantly solve new problems in new environments." said Brian Van Essen, a computer scientist at Lawrence Livermore National Laboratory, who is responsible for applying deep learning algorithms to national security.

TrueNorth chip is the latest technology crystallization. In the future, it will be responsible for the operation of deep learning and a series of other AI services. At present, the machines running Google, Facebook and Microsoft algorithms need to be equipped with independent graphics processors, but they are all moving towards FPGA (chips capable of programming specific tasks). Peter Diehl (Ph.D. from the cortical computing group of Zurich University of Technology) believes that TrueNorth is better than independent graphics chips and FPGAs because of its low power consumption.

Jason Mars, a professor of computer science at the University of Michigan, believes that the main difference is that the combination of TrueNorth and deep learning algorithm can be considered as seamless. Both simulated neural networks in depth, and generated neurons and synapses in the "brain". "The chip can efficiently execute the commands of the neural network." He did not participate in the trial run, but closely followed the progress of the chip.

Even so, TrueNorth is not fully synchronized with the deep learning algorithm. However, IBM has decided to let external researchers participate in the improvement of the chip, because it is still a long way from the real market. For Modha, this is also a process that must be experienced. As he said, "We need to lay a solid foundation for major changes."

Brain in mobile phone

Peter Diehl recently traveled to China, but for some reasons you know, his mobile phone can't use Google, which puts AI back to its original form. Because most cloud computing now depends on Google's servers, without the network, everything is in vain.

Deep learning requires very strong processing power. Generally speaking, this level of processing power can only be provided by giant data centers, while our mobile phones are generally connected to it through the network. TrueNorth is different. It can at least transfer some of its processing power to your mobile phone or other devices, which can greatly expand the use of AI.

But to understand this, you must first understand how deep learning works. Its operation is divided into two stages. First, companies like Google and Facebook need to cultivate their own neural networks to handle special tasks. If they want to automatically identify cat photos, they should first let the neural network see a bunch of cat photos. Then, after the pattern is trained, another neural network is required to perform this task. If you take out a picture, the system will judge whether there is a cat in it. The purpose of TrueNorth is to improve the efficiency of the second step.

Once the training of the neural network is completed, the chip can help you bypass the huge data center and directly complete the second step. And because the TrueNorth chip is small in size and power consumption, it can also be loaded onto handheld devices. This improves the overall efficiency, because you no longer need to download the calculation results from the data center through the network. If it can be popularized, it can greatly reduce the pressure on the data center. "This is the future of the industry. In the future, equipment can perform complex tasks independently," said Mars.

Neurons, axons, synapses and nerve impulses

Google has been trying to make neural networks enter mobile phones recently, but Diehl thinks that TrueNorth has a lot of advantages over other competitors. After all, it is more in tune with deep learning. Each chip can simulate millions of neurons, and these neurons can communicate with each other through "synapses in the brain".

This is what makes TrueNorth different from similar products in the market. Even compared with graphics processors and FPGAs, TrueNorth has sufficient advantages. TrueNorth chips can form "nerve impulses", similar to electrical impulses in the brain. Nerve impulses can show the change of tone in someone's speech, or the change of color in the image. "You can think of it as a small message passed between neurons." Rodrigo Alvarez Icaza said that he is one of the chief designers of chips.

Although the chip has 5.4 billion transistors, its power consumption is only 70 milliwatts. What about standard Intel processors? It has 1.2 billion transistors, but its power consumption is 35 to 140 watts. Even ARM chips commonly used in smart phones consume several times as much power as TrueNorth chips.

Of course, new software is needed to make this chip work, which is exactly what Diehl and other researchers have been trying to do during the trial run. In other words, the R&D personnel are converting the existing code into the chip's recognizable language and inputting it into the chip, but they are also working hard to write native code for TrueNorth.

Gifts

Like other researchers, Modha mainly focuses on discussing TrueNorth in the biological field, such as neurons, axons, synapses, and nerve impulses. There is no doubt that the chip simulates the human nervous system in some aspects, but it still has its limitations. "Such discussions often arouse people's vigilance. After all, silicon is not the substance that makes up the human brain." Chris Nicholson, co-founder of Skymind.

Modha also acknowledges these claims. When he started this project in 2008, he received an investment of 53.5 million dollars from Darpa (a research institution of the Ministry of Defense), with the goal of building a new chip with completely different materials and simulating the human brain. But he knew that this goal could not be achieved soon, and "we can not ignore the reality on the way to pursue our dreams," he said.

In 2010, he was bedridden due to swine flu. During this time, he realized that the best way to break the bottleneck was to start from the chip structure to simulate the brain. "You don't need nerve cells to simulate basic physics, chemistry and biology to improve computing power. We need to maintain enough flexibility while becoming more and more similar to the brain." He said.

This is the TrueNorth chip. It is not a digital brain, but it has taken an important step on this road, and through the trial run of IBM, this plan has also entered the formal stage. The whole machine is actually composed of 48 machines, each of which has its own TrueNorth processor. Next week, with the end of the trial run, Modha and his team will decompose the machine and let researchers take it home for further research. Human beings use technology to change society, and these researchers are our backbone.