Home > News content

"Tianji" on the cover of Nature: Shi Luping team released the world's first heterogeneous fusion brain chip!

via:博客园     time:2019/8/1 8:32:36     readed:531


Source: Nature, Editor: Daming, Zhang Jia, Xiao Qin

【新智元导读】good news! Tsinghua University has developed the world's first heterogeneous fusion brain chip, and posted the latest issue of Nature's cover! The research team also showed the "unmanned bicycles" driven by the chip. This research was led by Professor Shi Luping from the Brain Computing Research Center of Tsinghua University. It lasted for 7 years and finally became a positive result!

Tsinghua University has developed the world's first heterogeneous fusion brain computing chip & mdash; — "day movement", the chip-driven "unmanned bicycle" & rdquo; boarded the latest issue of Nature cover!


The study was conducted by a team of professors of the Brain Computing Research Center of Tsinghua University, based on the Department of Precision Instruments, to demonstrate a self-driving bicycle driven by a new artificial intelligence chip.


Paper based on the results of this research “Heterogeneous Celestial Chip Architecture for Artificial General Intelligence"Towards artificial general intelligence with hybrid Tianjic chip architecture" was published as a cover article on August 1st, "Nature".China's zero breakthrough in the two major fields of chip and artificial intelligence.


Celestial chip 5x5 array expansion board

At this stage, there are two main methods for developing artificial general intelligence:NeuroscienceBased on the human brain as much as possible; the other is based oncomputer scienceFor guidance, let the computer run a machine learning algorithm. Both have their own advantages and disadvantages, and the integration of the two is currently recognized as one of the best solutions. Developing a computing platform that combines the two will be a key to driving convergence. The new chip combines two technical routes, and this fusion technology is expected to enhance the capabilities of each system and promote the research and development of artificial general intelligence.

This hybrid chip is named“Ten movement”(Tianjic),There are several highly reconfigurable functional cores that support both machine learning algorithms and existing brain-like algorithms.

Researcher uses oneAutomatic drivingBicycle systemThe processing power of this hybrid chip was verified.

This is a heterogeneous and scalable manual universal intelligent development demonstration platform.Use a piece of heavenly chipIt demonstrates the functions of bicycle self-balancing, dynamic sensing, target detection, tracking, automatic obstacle avoidance, obstacle avoidance, speech understanding, and self-determination.

In the test, the unmanned bicycle can not onlyRecognize voice commands, achieve self-balancing control, detect and track forward pedestrians, and automatically avoid obstacles.


S-type route tracking


Voice control & ldquo; turn left & rdquo;


Voice Control & "Direct and Accelerate”


Autonomous obstacle avoidance

Professor Shi Luping said that this is only a very preliminary study, but this research can promote the further development of the artificial general intelligent computing platform.

Next, Xinzhiyuan brought a detailed interpretation of this breakthrough study and an interview with the Shiluping team.

I. "Day movement": AGI road that supports the convergence of computer science and brain-like computing

It is generally believed that there are two ways to implement Universal Artificial Intelligence (AGI): computer science orientation and neuroscience orientation. Due to the fundamental differences between the ideas, concepts and implementation schemes of these two roads, they rely on different development platforms and are incompatible with each other, which greatly hinders the development of AGI technology. There is an urgent need for a common platform that supports both approaches simultaneously.

Developed by the Shi Luping team“Ten movement”(TIanjic chip)This is done to provide a hybrid collaborative development platform for AGI technology.


Tianjic chip and test board

The Tianjic chip adopts a stream-like flow control mode of a multi-core architecture, a reconfigurable functional core module and a hybrid coding scheme. It can not only adapt to computer-based machine learning algorithms, but also easily implement neural computational models inspired by brain principles and various Coding scheme.


Telescope heterogeneous fusion brain computing architecture

With just one chip, you canSimultaneous processing of multiple algorithms and models in an unmanned bicycle system for real-time target detection, tracking, voice control, obstacle avoidance and balance control. This research is expected to open up new avenues for the development of more versatile hardware platforms and promote the development of AGI technologies.

Given the current advances in machine learning and neuroscience, AGI systems should have at least the following characteristics:

  • It can support the expression of rich spatial, temporal and spatiotemporal relationships in neural networks.
  • Support for layered, multi-granular and multi-domain network topologies, not limited to a specific network structure.
  • Support for various models, algorithms and coding schemes.
  • Support for interleaving cooperation of multiple specialized neural networks, which may be designed to handle different tasks in parallel.

These features need to run efficiently in a common platform that enables support for mainstream artificial neural networks (ANNs) and neuroscience-inspired models and algorithms in a unified framework.


Figure 1: A hybrid route to achieve AGI development

To support these capabilities, the team developed a cross-paradigm computing platform that can accommodate neural networks for computer science and neuroscience (Figure 1), compatible with a variety of neural models and algorithms, especially biologically based (such as pulsed neural networks) , ie SNN) elements.

In general, ANN and SNN have different modeling styles for information representation, computational principles, and memory organization (as shown in Figure 2a). The biggest difference between the two is that the ANN processes information with precise multi-bit values, while the SNN uses binary pulse sequences. An implementation comparison between ANN neurons and SNN neurons is shown in Figure 2b.

On the other hand, there are some similarities between ANN and SNN neurons, which leaves room for integration between models. By comparing the neural network models of ANN and SNN in detail, the computational model is parsed and mapped to the relevant neuron function modules - axons, synapses, dendrites and cell bodies - to construct a trans-paradigm unified neuron scheme. (as shown in Figure 2c). The team designed synapses and dendrites that applied both protocols, while axons and somatic cells changed function by independent refactoring.


Tianjic chip design schematic

Figure 2d is a complete schematic diagram of a single-function nuclear (FCore), including axons, synapses, dendrites, cell bodies, and routing. For deep integration, almost the entire FCore can be reconfigured to achieve high utilization in different modes. FCore covers the linear integral and nonlinear transformation operations used by most ANNs and SNNs. The FCores on the chip are arranged in a 2D 2D grid, as shown in Figures 2e and 2f.

The Tianjic chip and its back-end layout are shown in Figure 3a. The chip consists of 156 FCores and contains approximately 40,000 neurons and 10 million synapses. The Tianjic chip is fabricated on a 28nm semiconductor process with an area of ​​3.8× 3.8 square millimeters. The chip area occupied by each individual module, including the axon, current, signal, router, controller and other chip overhead, is shown in Figure 3b. Since resources can be reused, the areas used to be compatible with SNN and ANN models account for only about 3% of the total area. The power dissipation of FCore is shown in Figure 3c.


Summary of chip evaluation and modeling

Tianjic is capable of supporting a variety of neural network models, including neuroscience-based networks (such as SNN, and biologically inspired neural networks) and computer science-based networks (such as MLP, CNN, and RNN). Figure 3d shows the test results of testing different network models and general purpose processing units on the Tianjic chip.

As shown in Figure 3e, a hybrid neural network with dendritic relay can break through the limitations of the traditional neuromorphic chip Fan in/fan out, avoiding the loss of accuracy of the SNN network (+11.5%)..The added overhead of using this hybrid mode is negligible, because Tianjic can naturally implement heterogeneous conversions in FCore. Using Tianjic, you can also explore more biologically meaningful cognitive models (as shown in Figure 3f).

Second, voice control, automatic obstacle avoidance, this unmanned bicycle is very beautiful

In order to prove the feasibility of building a brain-cross-paradigm intelligent system, the team developed a heterogeneous and scalable manual general intelligent development display platform using unmanned bicycles, and deployed in parallel and running multiple private networks in a single Tianjic chip.

The bicycles in the experiment are equipped with a variety of algorithms and models to perform real-time object detection, tracking, voice command recognition, acceleration, deceleration, obstacle avoidance, control balance and decision making tasks (Figure 4a).


Automatic bicycle demonstration platform

To achieve these tasks, there are three main challenges to overcome:

First, in the outdoor natural environment, it successfully detects and smoothly tracks moving targets, crosses the speed bumps, and automatically avoids obstacles when necessary.

Second, real-time motor control signals are generated in real-time response to balance control, voice commands, and visual perception to keep the bike moving in the right direction.

Third, to achieve integrated processing and rapid decision making of a variety of information.


Unmanned bicycle test results based on Tianjic chip multi-model integration platform

To accomplish these tasks, the team developed several neural networks, including CNN for image processing and object detection, CANN for human target tracking, SNN for voice command recognition, and MLP for attitude balance and direction control. There is also a hybrid network for decision control.

Thanks to the chip's decentralized architecture and arbitrary routing topology, the Tianjic chip platform enables parallel operation of all neural network models and enables seamless communication between multiple models, enabling the bike to perform these tasks smoothly. Figure 4c shows the output signal in response to different voice commands.

Figure 4d shows the output control signal of the bicycle while tracking, obstacle avoidance, and "S-shaped" Figure 4e shows the learning of vehicle attitude and steering control at different speeds based on physical measurements.

The Tianjic chip can support both computer science-based machine learning algorithms and neuroscience-based biological models. It can freely integrate various neural networks and hybrid coding schemes to achieve seamless communication between multiple networks, including SNN and ANN.

In summary, this paper introduces a novel brain-like computing chip architecture that achieves flexibility and scalability by integrating cross-paradigm models and algorithms onto a single platform. It is hoped that this research will accelerate the development of AGI and promote the development of new practical applications.

Three or seven years of tempering "day movement", bicycle is a complete brain-like computing platform

For the questions that everyone is concerned about in this study, Shi Luping, a professor of precision instruments at Tsinghua University, and Yan Jing, an associate researcher at the Department of Precision Instruments at Tsinghua University, and Deng Lei, a postdoctoral fellow at the University of California, Santa Barbara, accepted the media interview. .

ask:What is the biggest challenge in the research?

Shi Luping:We have been cultivating this research since 2012 and have encountered many challenges, but we believe that the biggest challenge is not from science, nor from technology, but that the distribution of disciplines is not conducive to solving such a problem. We believe that multidisciplinary deep integration is the key to solving this problem. So in this study, we formed a multidisciplinary team composed of seven departments to form a brain-like computing research center covering brain science, computer, microelectronics, electronics, precision, automation, materials and so on. Here, I am especially grateful to the leaders of Tsinghua University for their strong support for interdisciplinary construction, which is the key to the success of this project.

Deng Lei:On the chip side, the biggest challenge is how to achieve deep and efficient integration. I emphasize two points:

the first, is deep and efficient. There are two types of neural network models that are relatively hot, one is from computer science, and the other is from brain science. The languages ​​of these two models are very different. They have different computing principles, different signal encoding methods, and different application scenarios, so the computing architecture and storage architecture they need are very different. Even if the design optimization goals are very different, this can be seen from some of the deep learning accelerators we can see today, as well as some neuromorphic chips, which are basically independent of the design system. Therefore, it can be seen that deep integration is not simple. It is not that designing a deep learning acceleration module, redesigning a neuromorphic module, and then putting them together. This is not feasible. It is difficult to determine each part. What is the ratio, because the application in reality is complex and changeable, which is not efficient.

secondIf you build a heterogeneous hybrid model, you may need to have a dedicated signal conversion unit between the two modules, which has a lot of extra cost, so how to design a chip architecture to be compatible with these two types of models, Moreover, it can be flexibly configured and has high performance, which is also a challenge in our chip design.

ask:Why choose an unmanned bicycle as an entry point?

Shi Luping:Bicycles are for our chips. At that time, we went through in-depth discussions to determine what kind of application platform to develop to show us a heterogeneous fusion of new features, which is not easy. We have four considerations:

First, we hope that this is a multi-modal system that is somewhat similar to the brain, rather than some algorithms like the current AI, only a single application. We hope that this is a complete link that covers perceived decision-making and execution, so that we can provide support for multiple models of heterogeneous fusion, so this is different from a single model.

Second, we hope that this is also able to interact with the real environment, rather than doing experiments in the computer room or doing a simulation in the computer. We hope that it can be a real environment interaction.

Third, we hope that this system is best for our processing chips with power consumption and real-time requirements, so as to reflect the advantages of our dedicated chips.

Fourth, because we have to do repeated experiments, we hope that this system is controllable and scalable.

By considering the above points, we finally chose the unmanned bicycle platform, which has the function of voice recognition, the function of target detection and tracking, motion control, obstacle avoidance and self-determination. So it looks small, but it's actually a small, brain-like computing platform.

ask:Can the brain like the brain?

Shi Luping:Everyone is interested in whether brain-like technology can transcend the human brain. In fact, this is the same as everyone is always asking how the computer will surpass the human brain.

The computer has long surpassed the human brain, just to say in what areas. All of us now think that the genius has the ability to be amazing. In fact, the current computer is very easy to implement, such as remembering fast, remembering, counting fast, calculating accurately, etc. In these respects, for computers, It is a pediatrics.

However, at present, at many levels of intelligence, especially for uncertainties, there are still considerable distances between computers and human brains in many fields such as learning and self-determination.

The computer will gradually narrow the gap. As for whether it can surpass the human brain in the end, I personally think that there will be more and more technical aspects, because the development of the computer has a characteristic that it never regresses, and it keeps moving forward. But I believe that we are wise. We will gradually improve our understanding of the field of research in the process of development, to control its risks, because I believe that people pay attention to this issue because we Worried about whether or not humans will be destroyed as sci-fi movies say.

In fact, what we can destroy humanity has already been created, that is, nuclear weapons, but why is it not destroying humanity now? Because we master it, we can control it. Like brain-like computing, strong artificial intelligence, and artificial general intelligence, we believe that humans can make good use of our wisdom to regulate its path of development so that it can benefit us and minimize those risks.

The cooperation units of this paper include Beijing Lingyi Technology Co., Ltd., Beijing Normal University, Singapore University of Technology and Design and the University of California, Santa Barbara.

Paper address:


China IT News APP

Download China IT News APP

Please rate this news

The average score will be displayed after you score.

Post comment

Do not see clearly? Click for a new code.

User comments