Shulou(Shulou.com)06/02 Report--

Produced by big data Digest

Author: big data Digest editorial Department

Today, a self-driving bike from Tsinghua University appeared on the cover of Nature.

The bike can not only balance itself, but also bypass obstacles and even respond to simple voice commands.

Bicycles can change direction or adjust speed according to sound commands.

Bicycles detect and track moving people and avoid obstacles if necessary

It depends on the brain behind the bike that the bike can run so smoothly and smoothly. It uses a new type of computer chip called Tianjic, which is used for real-time object detection, tracking, speech recognition, obstacle avoidance and balance control.

The team from Tsinghua also appeared on the cover of the latest issue of Nature released on Aug. 1 with its Tianjic chip.

This is also the first time that a Chinese artificial intelligence chip has been posted on Nature.

Professor Shi Luping, a communications author, professor of precision instruments at Tsinghua University and director of the brain-like Computing Center, said that although this is still a very preliminary study, it may be able to promote the further development of the general artificial intelligence (AGI) computing platform.

The authors of this paper are from Tsinghua University, Beijing Lingxi Science and Technology, Beijing normal University, Singapore University of Technology and the University of California, Santa Barbara.

Links to papers:

Https://www.nature.com/articles/s41586-019-1424-8

Computer science + neuroscience dual-oriented to build a more universal platform

There are two main ways to develop general artificial intelligence: one is neuroscience-oriented, rooted in neuroscience, and trying to build circuits that are very similar to the brain. The other is computer science-oriented, which is based on computer science and uses computers to implement machine learning algorithms.

Because of the basic differences in preparations and coding schemes, the two methods rely on different and incompatible platforms, delaying the development of AGI.

Therefore, the development of general artificial intelligence urgently needs a general platform to support more general artificial neural networks based on computer science and neuroscience-inspired models and algorithms.

The combination of the two methods promotes the development of AGI

This Tianjic chip integrates two methods to provide a hybrid and collaborative platform.

The Skymachine chip adopts multi-core architecture, reconfigurable construction module and streamlined data stream with hybrid coding scheme, which can not only adapt to the machine learning algorithm based on computer science, but also easily realize brain startup circuit and a variety of coding schemes.

Schematic diagram of "Tianji" chip

Through resource reuse, Tianji chip only needs 3% extra area to run most of the neural network models guided by computer science and neuroscience at the same time, support the hybrid modeling of heterogeneous networks, and form a time-space coordinated scheduling system. Giving full play to their respective advantages can not only reduce energy consumption, improve speed, but also maintain high accuracy.

The Skymachine chip has multiple functional cores at the same time, which can be easily reconfigured to adapt to machine learning algorithms and brain heuristic circuits. The researchers demonstrated the potential of this method by integrating one of the chips into a driverless bike, which is self-balancing, voice-controlled and can detect and avoid obstacles.

Example of chip evaluation modeling

In the paper, the team said that using only one chip, multiple algorithms and models can be processed simultaneously in a driverless bicycle system to achieve real-time object detection, tracking, voice control, obstacle avoidance and balance control.

Example diagram of multi-mode integration of self-driving bicycle based on Skymachine chip

Interdisciplinary team, with the participation of seven colleges and departments, sharpens a core in seven years.

At the telephone press conference on July 30, Shi Luping, the author of the paper and a professor of precision instruments at Tsinghua University, introduced the research ideas of the paper. In an interview with the media, the research team said that since the study was conceived in 2012, the biggest challenge faced by the team has not come from science or technology, but that the distribution of disciplines is not conducive to solving current problems.

Therefore, the research team set up a brain-like computing research center composed of seven departments, covering brain science, computer, microelectronics, electronics, precision instrument, automation, materials and other disciplines.

Deng Lei, a postdoctoral fellow at the University of California, Santa Barbara and a member of the team, said the biggest challenge in chips is how to achieve deep and efficient integration. Computer science-oriented and neuroscience-oriented are two kinds of popular neural network models at present. The languages, computing principles, signal coding methods and application scenarios of these two models are very different, so the required computing architecture and storage architecture are very different, and even the optimization objectives of the design are very different. Some deep learning accelerators and neuromorphological chips are basically independent design systems, so deep fusion is not simple.

Deep fusion is not a simple combination of deep learning acceleration module and neural morphology module, the difficulty is that the proportion of each part is difficult to confirm, because the application in reality is complex and changeable. Moreover, if you build a heterogeneous hybrid model, you may need to add a special signal conversion unit between the two modules, which will have a lot of additional costs, so how to design a chip architecture that is compatible with two types of models, can be flexibly configured and has high performance, is a major challenge in the team chip design.

In 2015, Shi Luping's team designed the first generation of "celestial movement". After continuous improvement, the second generation of "celestial movement" came out in 2017. Compared with the world's advanced IBM TrueNorth chips, the successful second-generation "heavenly movement" in 2017 has a density increase of 20%, a speed increase of at least 10 times, a bandwidth increase of at least 100 times, and better flexibility and expansibility.

As the MIT Technology Review reports, "the chip implies the progress China has made in developing its own chip design capabilities. Chinese researchers have shown that they can make specialized AI chips and any chip."

Can brain-like brains surpass the human brain?

In fact, as early as three years ago, Google released an ideal self-driving bike on April Fool's Day.

In Google's imagination, this "bike" is not only super balanced.

Can also automatically through the traffic lights, autonomous navigation to find your location.

However, the release as an "April Fool's Day video" also shows that this technology is difficult and difficult to achieve.

Today, the team of Tsinghua Shi Luping has finally realized this imagination. Such cool techs also seems to have a hint of science fiction, which makes people imagine the day when AGI arrives.

In an interview, Professor Shi Luping said that the question of whether the brain-like brain can surpass the human brain is actually similar to the question of whether the computer can surpass the human brain.

In fact, computers have long surpassed human beings in some respects, and we are amazed by their accurate and fast computing power and powerful memory. However, at present, at many levels of intelligence, there is still a considerable distance between the computer and the human brain. Especially for uncertain problems, such as learning, independent decision-making and other fields.

Computers will gradually narrow the gap, and as to whether they can surpass the human brain in an all-round way, Professor Shi Luping feels that there will be more and more from a technical point of view, "because the development of computers has a characteristic that it never goes backwards, it keeps going forward." But I believe that we are intelligent, and we will gradually improve our understanding of the field of research in the process of development, to control its risk, because I believe that people attach importance to this issue. Because we are worried about whether it will destroy human beings like science fiction movies say. "

On whether AGI will surpass human intelligence, Wu Enda also said in the AI For Everyone course that the emergence of a complete AGI may take decades or even hundreds of years, and in terms of time, we do not need to worry too much.

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.