When AlphaGo defeated Korean player Li Shishi, in order to better train his system, Google customized the "TPU" (Tensor Processing Unit) ASIC chip for machine learning, which is optimized for TensorFlow and has better performance. NVIDIA's GPU.
From the point of view of Google's emphasis on chips, the performance of the chip is very important for the artificial intelligence training itself, the process of machine learning. In the field of artificial intelligence, running deep neural network computing on CPUs and GPUs is nothing new. For the chip company, the future artificial intelligence will be more and more applied to various smart devices, and the requirements for the sensor and information processing speed of the chip are getting higher and higher.
In this way, scientists need to use specialized high-efficiency chips to deal with the massive data brought by deep neural networks. Now, using neural morphology calculations to simulate the technology of human brain processing information is becoming another technical direction in the field of artificial intelligence.
"The artificial intelligence deep neural network has a special computational structure, such as high degree of parallelization, recursion in the time domain, and sparse nodes in the middle. Therefore, if it can be implemented with special hardware, it will be more efficient than software implementation on the CPU. In general, it will increase by 2-3 orders of magnitude," Yu Kai, a Horizon robot company, said in an interview with the media.
Neuromorphic calculations can mimic human brain processing information
Von Neumann
Neuromorphic calculations, also known as brain stimulation calculations, have been an attractive target for scientists. The efficiency of the human brain has been an unattainable goal for many computers. In addition to being able to do more calculations with less energy, the most important thing is that the neuromorphic calculations are freed from the computational structure established by von Neumann, integrating the process of simulating brain processing and processing information onto the chip. In this way, machines equipped with such chips can learn data more quickly and efficiently.
However, most of the chips currently used in computers are based on the von Neumann architecture, relying on the central processor and memory to process information back and forth, and calculate the logic in the information. This method is great for working with numbers, executing well-written programs, but not for processing images or sounds. Take Google Inc., for example, when Google trains artificial intelligence to identify cats in video, it needs 16,000 processors to support it.
Neuromorphic calculations, on the other hand, hope to achieve an efficient effect by simulating the process of processing information in the human brain. It mimics the brain's billions of neurons and synapses to accept external information, such as vision, hearing, and subsequently, the information, pictures, and sounds that are received can change the connections between neurons. This whole process is the process of machine learning. In the neuromorphic calculation, a model similar to the human brain is also included, which can also be called a neural network.
IBM, Qualcomm layout neuromorphic computing chip
Neuromorphic calculations were first proposed by Carver Mead, a scientist at the California Institute of Technology who worked on brain-like research, in the late 1980s. Although the industry has been dominated by traditional chips in the past three decades, there are already technology giants working in this direction abroad, and even the United States Department of Defense Advanced Research Projects Agency (DARPA).
For example, IBM's TrueNorth project. Launched in August 2014, IBM designed a CMOS chip with neuromorphic engineering that includes 4096 hardware cores, each containing 256 programmable neuron chips with more than one million neurons. Synapses on neurons can accept signals and influence the connections between each other.
TrueNorth's project is one of the projects under DARPA (SyNAPSE), and the computing power of this chip is roughly equivalent to the brain power of rodents. It also bypasses the bottleneck of the von Neumann architecture and is also very energy efficient, with a power consumption of only 70 milliwatts and capable of operating 46 billion synapses per second.
If IBM's project is related to the military and is not suitable for commercial use, then the Qualcomm Zeroth project will have a stronger commercial taste. The Zeroth project not only hopes to mimic human-like perceptions, but also has the ability to learn how the biological brain works. Tony Lewis, senior director and product manager at Qualcomm's R&D department, thinks their research makes sense: "The brain is full of current activity and activity patterns such as consciousness, behavior, etc. People can't observe these activities before, and these patterns of activity are not going wrong. Knowing how to treat. With the results of Qualcomm's research, people will see the complex patterns of activity in the human brain and be able to identify them and develop methods for treatment recovery."
At present, Qualcomm has also used its own research results on toy robots, which allows robots to autonomously identify obstacles and actively bypass when they lose the signal transmission or disconnect the network. In the future, Qualcomm also hopes to establish a Zeroth platform and build more application projects.
The domestic level is not bad, and the road can be overtaken in the future.
The first 50 million neuron brain chips released by Xijing Technology
In addition to the investment of foreign technology companies, there are also start-up companies in China to join the research and development of neuromorphic computing chips. Shanghai Xijing Information Technology Co., Ltd. was established in 2015. In May of this year, the company launched a human brain real-time simulation simulator "Xijing Brain" with 10 billion "neurons", and also launched 50 million neuron brain commercial chips. Tan Limin, CEO of Xijing Information Technology, said in an interview with 澎湃News ("The use of cloud-free processing tasks in the absence of a network. We have previously tested accordingly, using our chips, within 8 seconds, 8 neurons The core has completed the learning of 60,000 pictures. The speed is much faster than the traditional CPU. It is precisely because of the way the human brain neurons work, which has the advantage."
According to Tan Limin, the company's unique algorithm can simulate the plasticity principle of human brain synapses, and also make the company a hardware team that can achieve "on-line learning". The company is still expanding its technical team and plans to launch mass-produced commercial chips by the end of this year. At the same time, the company also plans to use brain-like intelligence to help brain science research, and brain science research "reverse feeding" brain intelligence to cooperate with national brain planning team.
Although the chip of neuromorphic calculation has technical advantages, it has to be pointed out that from the current market situation, it is still dominated by traditional chips. Most companies are still exploring the application scenarios of the chip.
“Neuromorphic calculations are still in their infancy,†said Kathleen Schumann, a researcher at Oak Ridge National Laboratory. “We have not yet determined the specific architecture we will be running. We are also looking to simulate neuronal synapses in different ways. In addition, there are questions about how to actually use these devices, programming, etc. It takes time."
When Tan Limin was interviewed by the news, he also talked about some difficulties encountered in China. He told 澎湃News (): "In addition to the shortage of technical talents, what we lack is hardware resources. The chips and hardware we make involve the semiconductor industry. This industry itself needs to catch up with foreign countries. But in the research of artificial intelligence, The gap between China and foreign countries is actually not as big as imagined, but we started late, and the future 'curve overtaking' should be faster."
cheap usb flash drive,wholesale usb sticks,wholesale usb drives cheap, Modern cheap usb flash drives,bulk usb flash drives
Shenzhen Konchang Electronic Technology Co.,Ltd , https://www.konchang.com