Shulou(Shulou.com)11/24 Report--

This article comes from the official account of Wechat: back to Park (ID:fanpu2019), by Liu Kai and Jia Min

Today, with the development of self-driving technology, driverless vehicles have taken to the streets. However, in order to ensure driving safety, the vast majority of driverless vehicles are equipped with "safety guards" on the passenger seat to monitor the driving status of the vehicle. There is a joke: there is a driver hidden in the chassis of every driverless car. The leap from L3 to L4 to L5 seems to face an insurmountable graben.

If you want to keep pace with human drivers and be as flexible and independent as real people, there is another Plan B option for autopilot, and that is general artificial intelligence.

Written by Liu Kai (School of Educational Sciences, Bohai University), Jia Min (Institute of General artificial Intelligence, Bohai University)

Artificial intelligence has two different orientations: special artificial intelligence and general artificial intelligence, corresponding to special artificial intelligence and general artificial intelligence respectively. At present, the term "artificial intelligence" in academic and industrial circles almost without exception belongs to special artificial intelligence, which aims to solve specific problems through preset algorithms or training, while general artificial intelligence devotes itself to the research and development of meta-learning ability. with the help of nurturing to achieve the successful solution of specific problems.

Just as "intelligence" is regarded as "computing", autopilot seems to be a purely technical problem in the eyes of dedicated artificial intelligence experts, and tries to achieve technology expansion in a modular way. People can go out on their own without any equipment, but self-driving cars will not be able to move without the communication network. Even with the latest high-performance chips and lidar, self-driving cars often misjudge when driving off the highway or in a turn queue at slightly congested intersections. In order to solve such problems, automakers are scrambling to improve the intelligence of their cars, and the news of the rush from L2-L3 to L4-L5 is often reported. At the same time, all kinds of special computing equipment specifically to help self-driving vehicles cope with complex road conditions have emerged, and C-V2X cellular vehicle networking technology is one of them, which has functions such as intersection warning, accident warning, blind area monitoring, road emergency danger warning and so on. It is regarded as the key link of intelligent driving and intelligent transportation landing.

International Institute of Automata Engineers (SAE) rating of autopilot

Source: http://www.sae.org.cn/ articles / 18121901 predictably, the cost of self-driving vehicles passing through intersections will continue to rise, and what even experienced urban stray dogs can do, now have to deploy all kinds of monitoring equipment at all junctions, making the already complex urban road management even worse. This is limited to the city, after all, the "absolute safety" that can be obtained by external help is difficult to promote in the suburbs. It looks more like an urban "show" designed to be more "intelligent".

In the future, L4-L5-level autopilot under the route of special artificial intelligence technology will inevitably tend to the technical solution route of "intelligent outsourcing"-- not relying on stronger bicycle intelligence, but more complex and more sophisticated upper-level coordination nodes. There are many hidden dangers in this practice. the high-level coordination node not only has very limited coverage, but also will rapidly increase the marginal cost and overall complexity of the traffic system, prolong the bicycle decision chain and reduce the degree of decision participation. As a result, the systematic fragility of the intelligent road network itself will face the danger of being magnified at any time, which is the inherent weakness of the development route of dedicated artificial intelligence centralized technology.

In fact, the autopilot of dedicated artificial intelligence has never really faced the problem of "intelligence", but constantly transfers bicycle intelligence (internal solution) to seeking external assistance (external solution), that is, the vehicle itself is more responsible for raising problems than solving problems. In the end, it is not the cars that become more adaptable, but the ones that are artificially modified-the car itself is still sealed under the spell of "as much manpower, as much intelligence".

General artificial intelligence is different. It attempts to create a software system with cognitive function that can think and have emotion, and believes that intelligence is not omniscient, and the system needs to learn continuously through the preset meta-ability in order to reach the practical level in a certain field. Moreover, even after reaching this level, learning will not stop, especially in the context of changing circumstances. For general artificial intelligence, the innate preset is meta-level learning ability, but all the learning content is acquired. Therefore, the autopilot of general artificial intelligence system is essentially an educational rather than a technical problem.

Just as a newborn human baby cannot drive a car, the general artificial intelligence system is activated without any driving experience, let alone designed to drive a car. In fact, autopilot is a pseudo-problem for general artificial intelligence. Because, highly similar to human beings, the first step in the growth of general artificial intelligence systems is to gain direct physical experience through their own perceptual motion devices. There is no real difference between training a robot and training a car to complete the task of self-driving. Whether to assemble "organs" such as footsteps, tracks or wheels is only different in the "feeling" of the system, but not in "driving". Even, it is not necessary to install "legs and feet", and it is also possible to advance with wings or thrusters. Passengers can continue to sit in the car, or they can make the car in the shape of a robot and take passengers into their "bellies", or even carry them in their arms or on their backs. As long as it can achieve manned and purposeful autonomous movement, it is autopilot. Therefore, the autopilot of general artificial intelligence does not pick the type of equipment, the form of transportation and the driving environment. The autopilot of general artificial intelligence is the driving in the general sense.

On the one hand, this versatility is reflected in the diversity of peripheral choices, on the other hand, it is reflected in the breadth and understanding of the learning content behind driving. The whole meaning of the general artificial intelligence autopilot "algorithm" is to "raise" a general artificial intelligence "robot baby" from scratch. This "robot baby" is active, and all its actions depend on its own accumulated historical experience. It is an inward-dependent equal-power technology model, rather than an outward-dependent centralized technology model. Therefore, in the "nurturing" process of the general artificial intelligence system, there is no interference from the external God perspective of the special artificial intelligence, and the only God perspective is the internal perspective of the machine "baby" itself. The specific technical means of "parenting" are not the computer and robot-related software and hardware technologies such as machine learning, image recognition and optimization of special artificial intelligence, but the same educational means as human infants and young children. This particularity runs counter to people's understanding of machine learning and artificial intelligence over the past few decades, making it seem difficult to understand at first glance. Here, take perceptual motion as an example to illustrate:

First of all, the processing content of the special artificial intelligence system is all-inclusive, not only the concrete data such as image and sound, but also the abstract concepts such as natural language and knowledge. but the education of general artificial intelligence system can and can only start from perceiving movement experience. Because only direct experience is absolute experience, and only absolute experience can directly land in general artificial intelligence system. As for the idea of "imprinting ideas" on general artificial intelligence systems, it is a myth that these abstract experiences have no foundation at all, just as it is ridiculous to ask human babies that "good people are rewarded." Although in theory it may be possible to use some mysterious biological means to build specific neural structures in the baby's brain, in practice the baby has no understanding of this abstract idea. Therefore, even the most basic knowledge of autopilot, such as distance, speed, obstacle avoidance and even numbers, are unknown to robotic babies with general artificial intelligence at first, just like human babies.

Secondly, in the learning and training stage of special artificial intelligence system, perception and movement are usually separated. Image recognition does not require the camera to "jump" like the human eye, and the robot displacement is only the mechanical operation and optimization of the program. However, in the process of "nurturing" of the general artificial intelligence system, perception and movement are inseparable. They are not two different things, but different aspects of the same kind of things. Movement is the reason for the change of perceptual experience, and perception is the feedback of motion results. More importantly, perception can construct the intrinsic empirical meaning of the subject through movement directly or indirectly. For example, for a general artificial intelligence car with only wheels and ultrasonic sensors, the acoustic data itself is meaningless, and its meaning occurs after the motion changes because of it, and simple motion is meaningless. The significance of motion lies in the process of reducing the value of the sensor from it to zero. In other words, the perceptual motion signal is the most direct and inseparable atomic experience, which has no original meaning, and its meaning comes from the inherent mutual endowment. Take vision as an example, the active vision of general artificial intelligence is quite different from the basic principles and implementation of mainstream computer vision. Although they all "see", they have different "views". The former focuses on the meaning construction of top-down Affordance at motivation level, which is subjective and interpretable, while the latter focuses on bottom-up pattern discovery and matching at pixel level, which is essentially a mindless philosophical zombie.

In addition, the self-driving of general artificial intelligence is not only at the level of building cars, it is a mirror for us, which can reflect a deeper human truth and benefit us. For example, the movement of the general artificial intelligence machine "baby" can be divided into two categories: active movement and passive movement. The active movement of the wheel is controlled by the "baby" of the machine, while the passive motion is caused by the change of the position of the car body caused by the external force (such as the person picking up the car). Active motion is the engine of subjective experience, and it is the driving force to feel the self-boundary and distinguish the subject and object together with the sensor in the objective environment, so it is very important to the machine "baby". Only by experiencing such growth can we form a clearer identification between people and things, between ourselves and others, between ourselves and the environment. On the contrary, for those autistic children who are unable to effectively distinguish between people and things, self and others, the pathological enlightenment brought by general artificial intelligence autopilot, it is undoubtedly a glimmer of "alternative" dawn when the current autism biological hypothesis is in a quagmire.

Therefore, unlike today's drastic self-driving technology, general artificial intelligence is destined to follow a different path-not only driving vehicles, but even understanding "life".

Note:

The article is published in the Science and Humanities Section of China Social Science Daily.

Liu Kai, Jia Min. Explore autopilot based on general artificial intelligence. China Social Science Journal, 2022-08-23.

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.