Brain-like computing: let artificial intelligence go further

Thinking like a human being, this is what people expect from artificial intelligence and robots. The artificial intelligence that strode forward seems to have reached a crossroads.

"The comprehensive intelligence level of the machine is quite different from that of the human brain. Machine learning requires more manual intervention. There are fewer interactions between different artificial intelligence modes..." Recently, at the S43 Xiangshan Science Conference held in Hong Kong, the participating scientists Count the current bottlenecks in the development of artificial intelligence.

To solve the bell, you must also ring the bell, and the development of artificial intelligence is no exception. Scientists realize that if they want to go further, artificial intelligence will return to where it started, and that is human intelligence.

Neuroscience provides the foundation

"The machine learning method represented by deep learning is comparable to or even surpasses human level in specific problems such as audio-visual perception." In the conference report, as a neuroscientist, academician Pu Muming, a researcher at the Institute of Neuroscience of the Chinese Academy of Sciences, has been artificial intelligence in the past few decades. The development of the year is awesome. However, he also saw that compared with the learning ability of the human brain, machine learning has obvious gaps in terms of interpretability, reasoning ability, and ability to perform inferiority.

Letting machines learn from people is an important direction to improve the level of "smart". Academician Ye Yuru, Executive Chairman of the conference and Vice President of the Hong Kong University of Science and Technology, pointed out: "The goal is to theoretically simulate the mechanism and structure of the brain at multiple levels, and develop a more universal AI to deal with multi-task, self-learning and self-adaptation. The challenge."

“Brain inspired” is the most important development direction of artificial intelligence. In recent years, brain science research is transforming from the traditional understanding of the brain, the understanding of the brain to the protection of the brain and the enhancement of the brain and the process of the brain, that is, the transformation from "reading the brain" to "brain control" to "controlling the brain". Learning the brain's information processing mechanisms and building more powerful and versatile machine intelligence is very promising. The human brain working mechanism obtained through multidisciplinary crossover and experimental research is more reliable and is expected to provide a basis for the future development of artificial intelligence.

On the other hand, artificial intelligence can provide technical support for neuroscience and brain science in data collection, labeling and modeling to promote the development of brain science.

Breaking the "von Neumann architecture"

Universal-oriented artificial intelligence is inseparable from brain-like computing chips. Shi Luping, professor of the Department of Precision Instruments at Tsinghua University and director of the Brain Computing Center, said: "As a new technology developed from the way of human brain storage processing information, brain-like computing will be the cornerstone of artificial general intelligence."

Breaking the "von Neumann architecture" has become an important way to learn from the human brain information processing. It is understood that in the "von Neumann architecture", the computing module and the storage unit are separated from each other, and the CPU must first read data from the storage unit when executing the command. For each task, if there are ten steps, the CPU will read, execute, read and re-execute ten times in turn. Time and power consumption are spent on data reading, which limits the data processing capability. This is very different from the phenomenon that the brain processes a lot of outside information but has very low energy consumption.

Brain-like calculations are expected to make brain-like synapses on the chip. In May of this year, Huang Tiejun, a professor of the Department of Computer Science and Technology of Peking University, combined with a number of units to achieve a fine modeling of primate retinal nerve cells and neural circuits, and proposed a pulse coding model that simulates the retinal mechanism. Imitation retina chip.

"Retinal over-speed full-time vision chip uses nerve impulses to express visual information like a biological retina. The pulse emission frequency is 'overspeed'. The human eye is a hundred times more capable of 'seeing' the words of high-speed rotating blades. 'Full-time' refers to the nerves collected from the chip. The pulse sequence reconstructs the picture at any time.” Huang Tiejun said, “This is the basis for real machine vision. It is expected to reshape the visual information processing system and bring about changes in the fields of driverless, robotics and video surveillance. ”

However, the synaptic chip is still in the laboratory stage and has not yet become practical. Experts at the conference believe that brain-like computing is an exciting and daunting challenge.

Three hidden worries

The China AI Development Report 2018 shows that since 2013, the scale of investment and financing in the global and Chinese artificial intelligence industries has shown an upward trend. Experts attending the meeting noted that there are currently more than 4,000 companies related to artificial intelligence in China, but less than one-third of the companies that are favored or concerned by investors and willing to invest. Over-reliance on foreign ready-made source code, unclear commercial application path and scarcity of professional talents are the three major concerns of current artificial intelligence enterprises.

Since 2015, Google, Facebook, Amazon and other open source software for machine learning have led to the adoption of a large number of off-the-shelf source code. In the eyes of scientists, it seems that the advantage is lost at the starting line, and the craftsmanship is deep and refined in other people's systems. In this regard, we should focus on breaking through the basic fields, and strengthen the in-depth study of the underlying algorithm model represented by deep learning for the underlying technology of artificial intelligence.

As for the business application path is not clear, experts suggest that enterprises should not be too blind, should find the direction of force as soon as possible, AI project commercial application scenarios can be the key to its success or failure, quickly accumulate core technology advantages, create a business model, can make Products that really have market demand. At the same time, we should adhere to the development route of seeking truth from facts, and avoid the evolution of artificial intelligence into a "Great Leap Forward", overdraft research and industrial capital resources.