Marvin Minsky displeased with the development of artificial intelligence

Marvin Minsky is a famous American scientist, winner of the Turing Award in 1969, Medal of the Benjamin Franklin Institute in 2001, holder of patents for the main graphic display (1963) and confocal scanning microscope (1961). The book “Perceptrons” brought him fame in the 60s, which became the fundamental work for subsequent developments in the field of artificial neural networks.

In the 80s. Minsky and Goode showed how neural networks can self-replicate according to an arbitrary training program. Artificial intelligence can be cultivated in a very similar way to the development of the human brain, and intelligence can be the product of interactions with no intellectual parts (this theory is known as the “Society of Reason”).

The last lectures of an 80-year-old professor on artificial intelligence can be downloaded as a podcast in three parts: part 1 (21 min.), Part 2 (25 min.), Part 3 (13 min.). Below we briefly outline the main points that the guru touched.

Marvin Minsky notes that recently the influence of alternative approaches, such as neural networks and genetic algorithms (evolutionary selection), has significantly decreased in scientific research on the topic of artificial intelligence. Researchers returned to more traditional areas, but for thirty years have not made a significant step towards artificial intelligence of the so-called “common purpose”, which has common sense. Progress in this direction almost stopped in the mid-70s. No one should be deceived by the appearance of specialized programs with individual external signs of artificial intelligence, such as programs for optical character recognition, text reading, etc. Researchers use mathematical methods to develop AI, but these methods are practically unsuitable for creating "real" AI that can solve problems on the basis of experience.
')
The professor criticizes modern academic education, when students have been building and repairing robots for years - but there really is no science here. According to Minsky, about 40,000 students all over the world are engaged in these silly things. But serious experts on AI quite a bit.

Marvin Minsky talks about different aspects of developing AI, including the importance of gaming programs. He himself has long been fond of this. In 1957, with his participation, a program for playing checkers was created, and in 1963 she defeated the Connecticut state champion, while she had the opportunity to learn and adjust to the players (so much so that she started playing worse after playing with bad players). When the developers found the base of the games and loaded it into the car, it became almost invincible. Forty years have passed since then, super-supercomputers have appeared, and scientists in the field of artificial intelligence continue to improve various game programs. In addition, experiments began in the virtual worlds.

Virtual worlds are very important for the development of AI. The fact is that the real world is rather boring from the point of view of artificial intelligence because of its meaninglessness , whereas in the virtual space each object has a specific purpose.

Marvin Minsky is very critical of evolutionary methods of creating AI (genetic algorithms leave no record of the reasons why losers die during natural selection), and statistical methods (they are easy to implement and they work in self-study programs, but in the long run this is a dead end ). Nevertheless, the professor truly believes that computers may well have intelligence. “People have a lot of explanations for why computers can't think, but all these explanations are stupid,” says Marvin Minsky. For example, they say that a computer cannot know the meaning of its actions. But this is a philosophical problem, even if philosophers are looking for an answer. Many people also do not know the meaning of their actions. Even more ridiculous is the thesis of having a self-consciousness, because people themselves do not quite understand what self-consciousness is: this is one of the secrets of the brain.

To implement this AI, according to Minsk, you need to simultaneously use several methods for storing and processing information, including neural networks, mathematical logic (in rare situations for important dependencies), fuzzy logic and relational databases.

Many things can be borrowed from the construction of the human brain. For example, scientists recently discovered strange effects of “post-learning” in the brain. It turns out that a person's test results improve significantly on the second and third day after a training session, although there was no additional training, and the number of hours of sleep is a critical factor for improving results: six or eight. In general, there is no single center of control in the human brain: each problem is solved by relatively independent separate small sections (there are about 400 of them), which have up to 30 connections with others, and some are generally isolated. Scientists have much to learn about the principles of the human brain, and this information can be used to build AI.

However, at the same time, it is necessary to understand that people are accustomed to call “artificial intelligence” something that has not yet been created . As soon as we create something (be it object recognition or speech recognition), we immediately stop calling it AI.

All who are interested in research in the field of AI, Professor Marvin Minsky suggests reading his latest book, The Machine of Emotions , which was published last year. By the way, the drafts of all the chapters from this book are posted on the professor’s personal website .