You get a bonus - 1 coin for daily activity. Now you have 1 coin

Nonformalizability

Lecture



One of the most important and hard to challenge critical criticism of artificial intelligence as a sphere of human effort was formulated by Turing as an argument based on the "non-formalization of behavior". Essentially, this criticism comes down to the statement that human behavior is too complex to be described using any simple set of rules, and since computers are not capable of anything other than the fulfillment of many rules, they are not capable of show the same intellectual behavior as people. In artificial intelligence, the inability to express all that is required is called a specification problem as a set of logical rules.

The main supporters of these views were the philosophers Hubert Dreyfus, who wrote a number of influential critical articles against artificial intelligence, including "What Computers Can't Do" (What computers are not capable of doing) and "What Computers Still Can't Do", and Brother Stewart, with whom Hubert wrote the article "Mind Over Machine".

The state of affairs that these scientists criticized gained fame as “the good old artificial intelligence,” or abbreviated as GOFAI (Good Old-Fashioned AI); This term was suggested by Hoagland. In this case, it is assumed that GOFAI is based on the statement that all intellectual behavior can be represented using a system that forms logical reasoning based on a multitude of facts and rules describing the problem area under consideration. Therefore, GOFAI corresponds to the simplest logical agent. Dreyfus was right in asserting that logical agents really have a weak point, since they do not solve the problem of specification.

Probabilistic reasoning systems are more suitable for use in open problem areas. Therefore, the Dreyfus criticisms do not refer to computers as such, but rather to one particular way of programming them. However, it is quite possible to assume that a more correct title is for the Dreyfus article, “What is a First Order Logical Rule-Based System Without Learning Can't Do” (What systems are not able to do based on first-order logic rules that do not use learning tools) , it would be much less impressive.

According to Dreyfus’s views, human experience implies knowledge of certain rules, but only as a “holistic context” (or “foundation”) within which people operate. He gives an example of correct social behavior in the presentation and receipt of gifts: "Usually people, handing a gift suitable for the occasion, act within the framework of the circumstances in this case." Obviously, people have "an immediate understanding of how to act and what to expect." He makes the same statement in the context of a chess game: "A mid-level chess player may need to think about the next move, and the grandmaster simply sees that the position on the board itself requires a certain move ... the correct answer is in his head."

Of course, it cannot be denied that the main part of the thought processes of the person preparing the gift, or the grandmaster who chooses the move, is carried out at a level inaccessible for introspection on the part of an inquiring mind. But it does not follow from this that these thought processes themselves do not occur. An important question that Dreyfus does not answer is how the right move appears in the grandmaster's head. Let us remind the reader of one of Daniel Dennett’s comments below.

It seems that philosophers have taken on the role of interpreters of magic tricks. When they are asked how the magician puts his trick in sawing an assistant in half, they explain that there is nothing difficult in this: the magician does not actually saw anyone; he just makes people believe that he does it. If philosophers are asked: "But how does he manage to create such an impression?", They answer: "We are not competent in this."

The Dreyfus brothers' article proposed a five-step process of learning that begins with the processing of the information received based on the rules (carried out on the same principle as in GOFAI) and ends with the acquisition of the ability to instantly select the correct answers. But making their proposal, the Dreyfus brothers essentially moved from the category of critics of artificial intelligence to the category of its theorists - they proposed the neural network architecture, organized in the form of a huge "library of examples", pointing out several problems. Fortunately, all the problems they indicated were completely solved, some partially and others completely. The problems mentioned by the Dreyfus brothers are listed below.

  1. A qualitative generalization based on examples cannot be achieved without background knowledge. Dreyfusy argue that there is no way to use the background knowledge in the learning process of the neural network. But in reality, such methods have already been developed that allow the use of a priori knowledge in learning algorithms. Nevertheless, these methods are based on the fact that there is knowledge that is presented in an explicit form, and the Dreyfus brothers stubbornly deny this approach. According to the authors of this book, the views of these scientists are a good reason for a serious redesign of modern models of neural information processing, so that they can use the knowledge gained earlier in the learning process, on the same principle as such knowledge is used in other learning algorithms .
  2. Neural network training is one of the varieties of supervised learning, which requires early identification of relevant input data and correct output data. Therefore, according to the Dreyfus brothers, the neural network training system cannot operate independently, without the help of a human teacher. In fact, learning without a teacher can be done using methods of unsupervised learning and reinforcement learning.
  3. The performance of learning algorithms is reduced when using a large number of characteristics, and if only a subset of characteristics is selected, then, according to these scientists, "there is no way to introduce new characteristics if it is found that the current set does not allow for some facts that are digestible learning process. " In fact, new methods such as machines supporting vectors very successfully cope with large sets of characteristics. There is also a fundamental opportunity to develop new features, although this requires much more effort.
  4. The brain is able to direct its sensors to search for relevant information and process it to extract aspects relevant to the current situation. But Dreyfus says that "details of this mechanism are currently unknown and there are not even such hypotheses about his work that would direct artificial intelligence research along the right path." In fact, the area of ​​active vision based on the theory of information cost is devoted to the problem of choosing the correct orientation of sensors, and the theoretical results obtained have already been applied to create some robots.

In the end, many of the problems that the Dreyfus brothers focused on (background everyday knowledge, specification problem, uncertainty, training, compiled forms of decision-making tools, the importance of using agents that respond to the current situation, rather than disembodied inference machines), and the results achieved are embodied in standard projects of intelligent agents. According to the authors of this book, this is evidence of the progress of artificial intelligence, and not confirmation of the impossibility of the goal set before it.

Comments


To leave a comment
If you have any suggestion, idea, thanks or comment, feel free to write. We really value feedback and are glad to hear your opinion.
To reply

Approaches and directions for creating Artificial Intelligence

Terms: Approaches and directions for creating Artificial Intelligence