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

Author | Lu Shouqun, Honorary Chairman of China Open Source Software Promotion Alliance

Produced | AI Technology Camp (ID:rgznai100)

Editor's note: recently, there has been a debate in the industry about whether the potential of deep learning algorithms has reached the ceiling. Some people think that the application based on deep learning algorithm still has deep development space, while others think that the current key is to tap the reasoning and decision-making ability of artificial intelligence, which needs to transition from the perceptual stage to the cognitive stage. A hundred schools of thought contend from experts, but it also shows that it is time to explore a new stage in the development of artificial intelligence.

The discussion on "Evaluation of artificial Intelligence to a New stage" initiated by Lu Shouqun, Honorary Chairman of the China Open Source Software Promotion Alliance, has attracted extensive discussions from Chinese and foreign experts and other people in the industry. I hope that the content with thinking value will promote and inspire new breakthroughs in artificial intelligence. The discussion has been officially launched on the CSDN blog (see the address at the end of the article).

Readers are welcome to express their views on the development of artificial intelligence at the end of the article. We will send a technical book in the field of artificial intelligence for each of the three comments with unique opinions and the highest praise.

At present, the underlying theory of artificial intelligence is based on deep learning based on artificial neural network, while the application space of artificial intelligence supported by depth technology algorithm is mainly concentrated in (or limited to) image recognition and speech recognition. Artificial intelligence recognition is a comparison, when the information into the brain after the lack of processing, understanding, thinking, creativity and other steps, stay in perception, unable to achieve cognition.

Machine learning / deep learning algorithm emerged in the 1950s (it has been used ever since). Today, the potential of deep learning algorithm is close to the ceiling, which limits the expansion of artificial intelligence application innovation.

On the eve of May Day this year, Academician Xu Kuangdi asked, "how many mathematicians in China have devoted themselves to the research of basic algorithms?" Reflecting his concern about insufficient investment in basic research in domestic artificial intelligence, some mathematicians seemed to have overreacted at the forum, believing that in recent years, the red-hot artificial intelligence in China is wearing a magnificent and false veil. there is no breakthrough in the underlying theory, the core algorithm is absent, and the development of artificial intelligence is facing a "choke neck" dilemma.

In 2014, IBM studied brain-like algorithms and developed TrueNorth chips to support innovation in artificial intelligence applications. IBM develops TrueNorth chips for brain-like algorithms based on large-scale pulsed neural networks, which are made up of 4096 small computing kernels that form 1 million digital brain cells and 256 million neural pathways that work like "brain neurons" (unlike traditional artificial intelligence chips that run packaged instruction sequences).

"the world's first dual-control heterogeneous fusion brain chip" developed by Shi Luping's team of Tsinghua University was published on the cover of nature magazine in 2019, which is of great significance.

Developed by Shi Luping's team of Tsinghua University, published on the cover of nature magazine.

"the world's first dual-control heterogeneous fusion brain chip"

In 2014, the Shi Luping team of the brain-like Research Center of Tsinghua University developed brain-like technology, which integrated the brain-like computing algorithm based on pulsed neural network (SNN) and the deep learning algorithm based on artificial neural network (ANN) into a chip called TianJic, which reused resources and made use of cross-advantages to make artificial intelligence application innovation closer to the cognitive stage of "independent thinking".

The celestial movement should belong to the CGRA structure (this is a higher-level reconfigurable technology). The FCcore corresponding to Tianjic is a unified hardware architecture that combines the main algorithms of SNN and ANN, and supports both commercial applications and algorithm research on the same chip, which can be said to be the biggest innovation of Tianjic. For functional verification on self-driving bicycles, it should be said that Professor Shi's team chose to apply the scene in this way, which is eye-catching, attractive and powerful.

The second generation heavenly movement (2017) has the characteristics of high speed, high performance and low power consumption. It has more kinetic energy, better talkability and expansibility, and 10 times faster speed than the TrueNorth chip.

Now it seems that IBM research and development of True North chip, Intel research and development of Loihi chip, both focus on the underlying theoretical research, that is, focus on brain-like pulse neural networks and brain-like algorithms. Impulsive neural network (SNN) is a model that simulates the connection and operation of biological neurons, which generates neural electrical pulses for information transmission through calculation, which is very different from the weighted connection + activation of traditional networks. At present, academia and industry at home and abroad are devoting themselves to the research on SNN, hoping to break through the deep learning algorithm, but the research on the new algorithm of SNN is still in its infancy.

The published list of parameters carrying SNN chips is published as follows:

List of parameters of Pulsed Neuron (SNN) Chip

In recent years, artificial intelligence technology represented by deep learning algorithm has developed rapidly. In view of the fact that deep learning does not have independent thinking, creativity and inspiration as human brain activities, and deep learning has some defects, for the further development of artificial intelligence, people look forward to the emergence of new algorithms.

With the development of artificial intelligence, sprouts based on new algorithms have emerged on the basis of breaking through deep learning algorithms at home and abroad.

1. Based on the brain-like algorithm of biological pulse neural network, its research results have broken ground (see the examples of IBM, Intel, Tsinghua University and Zhejiang University).

2. The "brain-computer interface" algorithm connects human brain neurons with extra-brain deep learning robots (or manipulators, computers). For example, in August this year, the he Bin team at Carnegie Mellon University in the United States implanted a "brain-computer interface" chip into the human brain and successfully connected with brain neurons without invasion. from then on, human brain neurons can be used to control the machine based on human mind (thinking or imagination). A new "brain-computer interface" algorithm, developed by the Russian "brain-computer interface" company (Neurobotics) and the Moscow Institute of Physics and Technology (MIPT), published this year, uses a "brain-computer interface" to connect human brain (EEG) neurons to a deep learning network (in this case, a non-invasive electrode that does not need to be implanted into the brain), which is expected to be used to treat stroke patients. Facebook (Facebook) and the University of California, San Francisco (UCSF) released the "brain-computer interface" technology (published in the sub-journal Nature) in July this year, which can read human language in real time, type with ideas, and take ultra-high-precision cameras with human eyes.

3. Knowledge-driven core algorithm (deep learning algorithm is a pure data-driven algorithm). IBM Watson (Watson) studies knowledge-driven in medical artificial intelligence, establishes a large-scale knowledge base, studies knowledge representation and reasoning, focuses on the core content of artificial intelligence, and develops a core algorithm based on knowledge-driven or the combination of knowledge-driven and data-driven, which promotes artificial intelligence to rise from perceptual stage to cognitive stage.

To sum up, there are many possibilities for new algorithms to break through deep learning in the future:

Hardware implementation of Pulsed Neural Network and brain-like Intelligent algorithm

Knowledge representation or cognitive algorithm based on the combination of data and knowledge

A brain-computer interface algorithm for connecting the real brain with deep learning

Https://www.toutiao.com/i6765805835990860295/

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.