IBM made an artificial brain out of 48 chips

At its lab near SAN Jose, IBM has built an electronic rodent brain from 48 TrueNorth test chips, each of which can mimic a basic building block of the brain.

IBM made an artificial brain out of 48 chips

Under the leadership of the project leader Dharmendra Modha, we got up close and personal with the entire project. It is understood that its volume is like a bathroom medicine cabinet, which is covered with translucent plastic panels, and can clearly see the chips, circuit boards and colorful indicators inside it. It looks like something out of a 70s science fiction movie, but Modha says, "You're looking at a small rodent."

He's talking about the brain of a small rodent, or at least this stack of chips can fit into that brain. These chips act as neurons, the basic building blocks of the brain. Modha says the system can simulate 48 million nerve cells, roughly equal to the number of nerve cells in a small rodent brain.

At IBM, Modha ran the cognitive computing group, which invented the "neurochip." When he and his team first unveiled their invention, they used it for a three-week test run, supporting academics and government researchers at IBM's research and development lab in Silicon Valley. After connecting their own computers to the digital mouse brain, the researchers explored its structure and began writing programs for the TrueNorth chip.

Last month, some researchers had already seen this guy in Colorado, so they had programmed it to recognize photos and speech, and understand some natural language. The chip runs the "deep learning" algorithms that now dominate the Internet's artificial intelligence services, providing facial recognition for Facebook and real-time language translation for Microsoft's Skype. Still, IBM has a head start here because its research could reduce the need for space and power supplies. In the future, we may be able to put this artificial intelligence into mobile phones and other small devices, such as hearing AIDS and watches.

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

TrueNorth is the latest technology that will run deep learning and a range of other AI services in the future. Today's machines that run algorithms from Google, Facebook and Microsoft still require separate graphics processors, but they are all moving towards FPgas (chips that can be programmed for specific tasks). Peter Diehl (PhD in the Cortex Computing Group at Polytechnic University Zurich) believes TrueNorth is superior to both standalone graphics chips and FPgas because of its low power consumption.

The main difference, says Jason Mars, a professor of computer science at the University of Michigan, is that TrueNorth works seamlessly with deep learning algorithms. Both simulate neural networks in depth and generate neurons and synapses "in the brain." "The chip can efficiently execute the commands of the neural network." He did not participate in the test run, but has closely followed the chip's progress.

Even so, TrueNorth is not yet fully synchronized with deep learning algorithms. However, IBM has decided to involve outside researchers in improving the chip, because it is still some distance from the actual market. For Modha, it was also a necessary process, as he said: "We needed to lay a solid foundation for a major transformation."

The brain in the phone

Peter Diehl recently traveled to China, but for some reason you know, his phone didn't work with Google, and he suddenly took artificial intelligence back to its original form. Because most of the cloud computing now depends on Google's servers, so without the network, everything is useless.

Deep learning requires a tremendous amount of processing power, which is typically provided by giant data centers, and our phones are usually connected to them via the Internet. TrueNorth, on the other hand, can move at least some of its processing power to your phone or other device, which could greatly expand the frequency of AI use.

But to understand this, you first need to understand how deep learning works. It works in two stages. First, companies like Google and Facebook need to build their own neural networks to handle specific tasks. If they want the ability to automatically recognize cat photos, they have to show the neural network a bunch of cat photos. Then, after the pattern is trained, another neural network needs to perform this task. When you take out a photo, the system has to determine if there are cats in it, and TrueNorth exists to make the second step more efficient.

Once you've trained the neural network, the chip can help you bypass the giant data center and go straight to the second step. And because TrueNorth's chip is so small and power efficient, it can fit into handheld devices. This increases overall efficiency because you no longer need to download the results from the data center over the network. If it can be popularized, it can greatly reduce the pressure on data centers. "This is the future of the industry, where devices can perform complex tasks independently." "Mars said.

Neurons, axons, synapses and nerve impulses

Google has recently been trying to bring neural networks to mobile phones, but Diehl thinks TrueNorth is way ahead of its rivals, because it's more in sync 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 sets TrueNorth apart from similar products on the market, even compared to graphics processors and FPgas have enough advantages. TrueNorth chips can form "nerve impulses," similar to electrical impulses in the brain. Nerve impulses can show a change in tone in someone's speech, or a change in color in an image. "You can think of it as little messages between neurons." Rodrigo Alvarez-Icaza, one of the chip's lead designers.

Although there are 5.4 billion transistors on the chip, its power consumption is only 70 milliwatts. What about standard Intel processors? It has 1.2 billion transistors, but its power consumption reaches 35 to 140 watts. Even ARM chips, which are commonly used in smartphones, consume several times more power than TrueNorth chips.

Of course, for the chip to really work, it needs new software, which is exactly what Diehl and other developers have been trying to do during the test run. In other words, the developers are converting existing code into a language that the chip recognizes and feeds into it, but they're also working on writing native code for TrueNorth.

present

Like other developers, Modha focuses on discussing TrueNorth in the field of biology, such as neurons, axons, synapses, nerve impulses, etc. The chip undoubtedly mimics the human nervous system in some ways, but it still has its limitations. "These kinds of discussions are often very cautionary. After all, silicon is not what the human brain is made of." Chris Nicholson, co-founder of a company called Skymind.

Modha acknowledges these claims. When he started the project in 2008, with a $53.5 million investment from Darpa (the research arm of the Department of Defense), the goal was to build an entirely new chip out of completely different materials and simulate the human brain. But he knows it won't happen quickly, and "we can't ignore reality on the way to pursuing our dreams," he said.

In 2010, he was bedridden with swine flu, during which time he realized that the best way to break through the bottleneck was to start with the chip structure and achieve a simulation of the brain. "You don't need nerve cells to mimic basic physics, chemistry and biology to improve computing power. We need to be flexible enough to become more and more like the brain." 'he said.

This is the TrueNorth chip. It's not a digital brain, but it's an important step along the way, and with IBM's trial run, the plan is on track. The entire machine is actually made up of 48 separate machines, each with its own TrueNorth processor. Next week, with the trial run over, Modha and his team will break down the machine for researchers to take home for further study. Humans use technology to change society, and these researchers are the backbone of our efforts.