"Human-Level Artificial Intelligence? Be Serious!" Read the papers report
First, the paper basic information
Author : Nils J. Nilsson. Niels Nielsen is professor at Stanford University, in 1958 received a doctorate in electrical engineering from Stanford University, more than 20 years at the Stanford Research Institute International Advisory Center for Artificial Intelligence. In addition to teaching artificial intelligence and machine learning, Professor Nelson also carried out research of flexible robot can respond to dynamic in the world, plan a course of action and learning from experience.
Publication Source : "Ai Magazine", 2005, 26 (4): 68-75.
Summary : I think the real human-level artificial intelligence is that most human work can be automated. I do not build special automation systems to achieve this goal, but to develop a common, acceptable education system, the system can be learned and taught to perform any one of thousands of human beings can work in the office. Combined with similar suggestions made by others, I have advocated the construction of a miniature, extensive capabilities and built-in-one system began. These must have the ability to self-learning and other capabilities.
Second, the paper details
This article mainly from artificial intelligence development objective, tells the story of how it should develop artificial intelligence, focusing on the development aspects of development and the challenges encountered in the process, finally gives the views of conclusive: the same as human intelligence system will eventually be realized.
Thesis can be divided into six parts:
AI to work
This part focuses on the goal of artificial intelligence to develop, the authors believe that the only machine to pass the Turing test does not constitute an appropriate or useful standard architecture of intelligent human beings. The author proposes a concept of "employment test", and use it instead of the Turing test. Employment test can detect whether they have the ability to complete the human machine to complete the work. Some people think, so that machine intelligence to achieve the level of human intelligence too difficult. Many researchers tend to study "weak AI", that is more focused on helping humans rather than replace human work to work. Author objected that true artificial intelligence is to achieve human-like intelligence, like intelligence, and that it has a positive meaning human society.
challenge
This section focuses on artificial intelligence research challenges encountered. If it wants to test the machine intelligence, machine intelligence will test whether there is the ability to achieve self-learning. Because some school sites and have the appropriate courses and tests to measure progress and ability teaching skills, so the author as a platform, the ability to judge AI program passed the exam, and thus see the results as the basis, if this skill enhancement, proof AI in moving to human intelligence. Of course, the authors do not recommend a specific skill tests or school website, as this is likely to be derived from a lack of ability to adapt to changes in the system, although it increases the efficiency of people.
Habile system
In this paper, the author, we ignore human beings are born with different talents and different facts "intelligence" of the inherent ability, but we still believe that human beings can learn a variety of different skills. Similarly, the authors believe that AI should strive to build a small number of general-purpose machines that are capable of self-learning and learning skills, instead of writing a larger number of machines, each machine has a different personal skills, then these skill sets from scratch. In the early essay, the author of this versatile machine called "habile systems" (Nilsson 1995). To achieve the main objective of humane AI, you can create several sub-goals to learn vocational skills in artificial intelligence projects. Sub-goal is not easier than the main objective, but at least we only need to build several such systems, rather than thousands of different work to build a single system. This sub-study will involve a major shift in the long-term goal of artificial intelligence research directions.
Children's Machine
Authors believe machine intelligence should be as intelligent as humans, to grow from childhood to adulthood. Start with the simplest and most easy to implement intelligent children start, the machine intelligence to imitate children's intelligence, then slowly develop. In this paper, Turing think we can not expect children to find a good machine at the first attempt, artificial intelligence has been carrying out years of experiments. The author wants to propose a core service, namely AI. Appropriate system with self-learning ability to provide the best way for human resource management. The authors believe that the original core indeed be quite complex. In addition, core and subsequent stratification must "receive proper education."
Core recommendations
Sensing system
core sensing system should include at least similar to typical human infant or child has similar capabilities. Enter the system should include visual, tactile, audio and tactile sensing mechanism. Must provide a basic set of procedures perception, it may be affected by a set of built targets, in order to deal with these inputs. It should include actuating motor output, sound, visual display and motion. A basic program activation respond appropriately to the sensing system and the target, provide an initial means for controlling these outputs. Early experience in a suitable environment for learning system provides the opportunity to increase these built-in capabilities.
Perception, representation and hierarchy of actions
the core system should have represent abstract hierarchical relationship model (using basic learning and perception predicates glossary) hierarchy and motor routine means. Author think of James Albus (1997) real-time control system (RCS), blackboard system (Engel-more and Morgan 1988) and the author of three towers architecture (Nilsson 2001). Systems need to be able to automatically generate such a hierarchy based on experience and education.
Forecasting and planning
core systems need to be able to predict the future state of perception, these proposed measures will be taken under the state of the current state of production. These forecasts form the basis to develop an action plan to achieve the target. The difference between the predicted value of the award of state and the state actually encountered is the basis of spatial and temporal difference learning method, (Sutton 1988) and animal (Montague, Dayan and Sejnowski 1996) prediction accuracy. About plan, McCarthy (1999) speculated that babies have the inherent ability to recognize that a prerequisite for action to achieve their goals should be as a child to pursue goals. He calls it "the target return." Target return is the basis for many AI planning system.
Learn
The core essence of the system is that it is through learning, imitation, experience, practice and education to growth and ability to change, at a convenient time and enhanced by reprogramming. At some stage the development of core systems, the enhanced perception, characterization, and mechanisms of action levels must be able to learn with the previously mentioned (Barto and Mahadevan, 2003) describes the relationship between the state and the (Dzeroski, De Raedt and Driessens 2001) linked. These are still important areas for future research.
Reasoning and knowledge representation
reasoning and use of knowledge representation attracted attention. AI research since its inception, we clearly use logical methods of reasoning is sound very difficult, if reasoning is a large number of statements from the beginning of the set, these statements you need a competent person must have the appropriate knowledge. However, many tasks require effective human reasoning, a strategy is to solve an unrealistic large knowledge base into two semi-independent part of the weak interaction.
Language
language skills needed for a variety of occasions: in between different people communicate, communicate with people, learn through reading, as well as internal representation and reasoning. Of course, also take the course and exam! By clear information or advice given command language, it gives a more accurate direction than reward and punishment implied.
in conclusion
The author believes that as humans intelligence system will eventually be realized. In any case, people will continue to try to automate human work, because the hardware and software implementation of their final salary and benefits than human cheaper. Even if the work due to economic, social or other reasons do not have the ability to become automated, we can still make the candidate system performance tests to measure these competencies required to work in order to measure the progress of human AI.
Third, the paper reading experience
This paper is a review of the paper, after reading this paper, we develop the goal of artificial intelligence, direction, methods and challenges with in-depth understanding. We believe that the ultimate beneficiaries of artificial intelligence or human. Because we have artificial intelligence, our life has become more interesting and convenient. Currently artificial intelligence has created mankind out of the very considerable economic benefits, artificial intelligence can replace humans do a lot of humans do not want to do, can not do the work, and the probability is lower than people make mistakes machines, and can continue to work, greatly enhance the work efficiency, saving a lot of cost, the future of artificial intelligence may also replace human work, replace humans do housework, help people learn, and even care for the elderly and children, health guardian class of real-time, direct human for treatment of sick and prolong human life, so that human life has become more and more beautiful.
