On September 1st, a press conference on the major achievements of billion-level neuron-type brain computers was held in Hangzhou. Academician Wu Zhaoxuan, President of Zhejiang University, attended and spoke. He said that the wave of artificial intelligence is accelerating the era of intelligent enhancement, brain-like computers will become the main form and important platform for future computing, will simulate brain function, efficient implementation of AI algorithms, improve computing power and other important unique role.
Facing the future, interdisciplinary gathering will become a new method to solve major problems, and system innovation based on multidisciplinary and multi-disciplinary fields will become an effective form of developing brain-like computers. It is hoped that today’s innovation of Zhejiang University and Jiangjiang Laboratory will be a big step forward in the development of a better life for mankind.
Academician Wu Zhaoxuan, President of Zhejiang University, attended and spoke.
Zhu Shiqiang, director of the Jiang Laboratory and deputy secretary of the Party Committee of Zhejiang University, said that the scientific research teams of the two sides have completed the research and development design day and night, and this stage of achievement is a major milestone. In the future, the project team will develop a larger-scale neuron brain computer based on China’s independent property rights brain chip, and study the basic software system of brain type supporting its operation and development, and gradually realize open source and openness, and contribute to the development of new brain computing technology in China.
Zhu Shiqiang, director of The Jiang Laboratory and deputy secretary of the Party Committee of Zhejiang University, delivered a speech.
The scene of the press conference on major achievements.
The three standard cabinets, 1.6 meters high, stand side by side, the black shell gives a cool feel, the red signal lights keep flashing, and closer seems to hear the sound of the pulsed signal running fast inside.
Recently, Zhejiang University jointly developed the first brain-like computer (Darwin Mouse) based on independent intellectual property rights brain chips in China.
This brain-like computer contains 792 Darwin 2 generation brain chips developed by Zhejiang University, supporting 120 million pulsed neurons, nearly 100 billion synapses, and the size of the number of neurons in the mouse brain, typical operating power consumption only 350-500 watts, and it is currently the largest neuron-like brain computer in the world.
At the same time, the team also developed a special brain-like computer-oriented operating system – Darwin-like brain operating system (DarwinOS), to achieve effective management and scheduling of brain-like computer hardware resources, to support the operation and application of brain-like computers.
Disrupt the traditional new computing model.
For computers that are already commonplace in all areas of work and life today, perhaps you’ve forgotten that scientists wanted to simulate a human brain by machine in the first place.
However, the development of computers, at that time chose to numerically long von Neumann architecture, that is, in the way of numbers plus subtranscence and subtranslation of information architecture. With the gradual failure of the Moore theories, the limitations of von Neumann’s architecture are becoming more and more obvious, and the problems of storage walls, power walls, and intelligent upgrading make the current computer development face major challenges.
For example, the storage wall problem is due to the existing Von Neumann architecture data storage and calculation of the separation, “it is like information stored in A, to calculate the time to move the information to B, calculated and then moved back to A.T. However, the speed of handling is much lower than the calculated speed, but let the handling itself become a key bottleneck. Pan Gang, head of the research team and a professor at Zhejiang University’s School of Computer Science and Technology, said the computing model constrained the performance of computing, represented by big data. The resulting data “run”, as well as high energy-consuming computing such as artificial intelligence, has led to power wall problems. At the same time, data-driven intelligent algorithms, training needs a large number of samples and intensive calculations, but to cite three, self-learning and other advanced ability is relatively poor, “now the machine intelligence is far from human intelligence.” “
The researchers are discussing.
How to break through the computer bottleneck caused by the way the existing computing works?
Scientists around the world are once again focusing on the original dream of mimicking the biological brain, developing new computing techniques by simulating the structure and computing mechanisms of the human brain in order to achieve energy efficiency and intelligent computing.
The biological brain naturally produces different intelligent behaviors in the process of interacting with the environment, including speech understanding, visual recognition, decision-making tasks, operational control, etc., and consumes very low energy. In nature, many neurons well below 1 million insects can do real-time target tracking, path planning, navigation, and obstacle avoidance.
Brain-like computer application demonstration: olfactory recognition.
Pan Gang said that using hardware and software to simulate the structure and operation mechanism of the brain neural network, the construction of a new artificial intelligence system, this subvert the traditional computing architecture of the new computing model, is brain-like computing. It is characterized by the one-in-one, event-driven, highly parallel, etc. , is the focus of international academic and industrial research, but also an important scientific and technological strategy, “brain-like computing has been regarded as one of the important paths to solve the computational problems such as artificial intelligence.” “
In recent years, Zhejiang University has focused on the core areas of human intelligence and machine intelligence, implemented the brain science and artificial intelligence fusion research plan referred to as the “Double Brain Project”, hoping to learn from the structure model and functional mechanism of the brain, apply the cutting-edge achievements of brain science to artificial intelligence and other research fields, and establish a new computer architecture leading the future.
In 2015 and 2019, Zhejiang University successfully developed Darwin 1st generation and Darwin 2nd generation brain computing chips, using chips to simulate the structure and functional mechanism of brain neural networks, which have advantages in the obfuscation of images, video and natural language. The result of this time is the integration of 792 Darwin 2 generation brain computing chips with our own property rights into three 1.6-meter-high standard server chassis, forming a powerful rack-type brain computer.
So how is this high-efficiency, low-power power achieved? Associate Professor Mader, the backbone of the project, said that the working mechanism of brain neurons is that the flow of potassium ion sodium ions causes changes in cell membrane voltage, thus transmitting information, “can be simply understood as a neuron receiving input pulses, resulting in an increase in membrane voltage in the cell body, when the membrane voltage reaches a certain threshold, will emit an output pulse to the axon, and through synapses to subsequent neurons to change their membrane voltage, to achieve the transmission of information.” “
It is very important here that the asynchronous operation, that is, when the signal comes to start, no signal does not run. Brain-like chips work in a similar way to the neuron behavior of living organisms, transmitting signals through pulses, which enables a high degree of parallelism and increased efficiency.
A demonstration of brain-computer interaction.
Really “thinking” like a brain.
With hardware, you have to have software.
Project research backbone Jin Xiaofei introduced, each chip has 150,000 neurons, each 4 chips made into a board, a number of boards and then connected to become a module. This kind of brain computer is like building blocks.
It’s easier said than done, but it’s not so simple to have so many neurons connected and expanded to enable efficient linkage combinations, while at the same time orderly assigning cluttered information flows to the corresponding functional brain regions.
To do this, researchers have developed a brain-like operating system for brain-like computers, DarwinOS.
This Darwinian brain operating system, which is a hybrid computing architecture for von Neumann architecture and neuromorphic architecture, realizes the unified scheduling and management of heterogeneous computing resources and provides a running and service platform for large-scale pulsed neural network computing tasks. Project research backbone Lu Pan said: “At present, Darwin-class brain operating system functional task switching time up to microseconds, can support hundreds of millions of brain-like hardware resource management.” “
As a result, the value of brain-like computer research can be realized — either in intelligent task processing in life or in neuroscience research, providing neuroscient scientists with faster, larger simulation tools and new experimental tools to explore the workings of the brain.
Demonstration of brain-like computer applications: Mind typing.
At present, researchers at Zhejiang University and its Jiang Laboratory have realized a variety of intelligent tasks based on Darwin Mouse-like brain computers. The researchers used brain-like computers as intelligent centers to realize the collaborative work of many robots in flood rescue scenarios, involving speech recognition, target detection, path planning and other intelligent tasks at the same time, as well as coordination between robots. At the same time, a number of different brain regions are simulated by brain-like computers, a neural network model of the outer knee-like nucleus of the thale brain is established, and periodic reactions of neurons in the brain region are simulated when the visual stimulation of different frequencies flashes;
Brain-like computer application demonstration: multi-robot coordination flood rescue.
Reporters at the experimental site saw that three similar-shaped robots, after simple training, to cooperate in flood relief missions. See the No. 1 robot with its own camera began to patrol the site, when found the dam gap, called the project responsible for the No. 3 robot to repair the dam, while searching for the injured, when found on the ground after the mannequin, and called the rescue robot No. 2. Robot No. 3 and Robot 2 came on a mission, and Robot No. 1 went on patrol somewhere else.
Brain-like computer application demonstration: multi-robot coordination flood rescue.
This scene doesn’t seem new, and existing robots can do it. But the biggest difference is that these robots are controlled by brain-like computers to carry out mobile commands by voice and accept task assignments. “The tasks of different robots can be switched by instruction, which means that their functions are not fixed, but are controlled by different brain regions, and robot No. 1 now does patrol work, which can then be turned into rescue or engineering.” Associate Professor Li Ying, the backbone of the project, said.
In another experimental scenario, members of the task force sing two sentences of a song to a computer, which can then “sing” the content of subsequent songs by going back.
“This is a brain-like computer that simulates the memory mechanism of the sea mass, enabling access to memory information inside the brain, unlike our usual search functions.” Professor Tang Huajin, the backbone of the project, said that Darwin Mouse-like brain computers can simulate the memory-learning function of the sea mass by drawing on the network structure of the seaum and the neural mechanism, and by coding the pulse of memory, the same model can learn and remember different types of data such as speech, songs and text.
Brain-like computer application demonstration: simulating memory recall of the sea mass.
How brain-like computers will “evolve”
The world’s first computer, born in 1946, weighed 28 tons and had a computational speed of 5,000 additions per second, but over the next 70 years, computer technology developed rapidly. The rate at which brain-like computers are developing is likely to be surprising.
Don’t look at today’s brain-like computers as a “big man”, scientists say, as Darwin’s chips and other hardware continue to iteratively upgrade, shrinking in size will be available. In the future, brain-like computers may be implanted into mobile phones and robots to create new intelligent service experiences.
How to make brain-like computers smarter than hardware updates is the focus of scientists’ next steps.
At present, the sensor input signal on the market is still mainly digital, in the application to Darwin Mouse brain computer, to add a coding layer, the signal into a pulse, and in the process, information loss and damage, will reduce the effectiveness of the computer to a certain extent. If this problem can be solved, brain-like computers can become smarter.
At present, brain-like computing research is still in its infancy, Darwin Mouse brain-like computers, both in terms of scale and intelligence, and the real human brain is still a big gap, but its significance is to provide an important practical example of this technology path, to provide researchers with a tool and platform to verify brain-like algorithms, with greater robustness, real-time and intelligent to solve practical tasks.
Take a photo with the research team.
The goal of researchers at Zhejiang University and Jiangjiang Laboratory is to one day make brain-like computers as versatile as von Neumann architecture computers, truly working as efficiently as the brain, and co-existing and complementary with von Neumann’s architecture to solve different problems as neuroscience and the maturity of system software, tool chains and algorithms for brain-like computers.
One industry insider said it was an important change in computational patterns, from numerical calculations such as multiplication and subtranscing to simulating the brain’s pulse calculations. “We want to be able to continue to move Darwinian brain-like computers in the direction of human intelligence, like biological evolution, to provide stronger artificial intelligence at ultra-low power consumption,” Pan said. “
Text reporter: Ke Yuen-jin Wu Yalan.
Photojournalist: Lu Shaoqing.