R. Kurzweil - How to Create a Mind (“How to create a mind”). Short review of the book

Ray Kurzweil is a rather interesting person. To begin at least with the fact that he was the first to invent musical synthesizers in 1965. He once received a whole series of awards for his inventions in various fields, published several books (The age of intelligent machines - 1990, The 10% solution for a healthy life -1993, the age of spiritual machines - 1998, The singularity is near - 2005), shot the film and even founded (along with Google and NASA) Singularity University. So when there was information that he was releasing a new book on how to make the brain (I thought it revealed the secrets of human thought), I immediately made a pre-order.



And for those who doubt whether it is worth buying and for all those interested in, I offer a sort of review: what exactly did the author want to say?

')





The book has 11 chapters + epilogue with only 321 pages, so for those who are familiar with English (and Ray loves and knows English :)), it will not take long to get acquainted with the book, it is quite possible to master the weekend.



When I did pre-order, I was almost sure that I would get a manual, where it would be described in detail how to actually make the mind and take over the world. The reality, as usual, has introduced its own corrections. It turned out that the book basically presents a sort of insight into the history of philosophy in general and of computer technology in particular. Actually, the methods used by Ray in his projects (such as, for example, a speech recognition project) are allotted only one chapter. Before that - basically the story, how they came to these methods, after that - why these methods should be perceived as a recipe for creating a mind.



So let's go through the chapters: the introduction and the first two chapters (“Mental experiments on the world” and “Mental experiments on thinking”) in the best traditions of “beginnings from afar” tell about how human thought developed and what inventions were made on the tip of a pen. History lovers will be interested, for everyone else - informative. The truth is already in the second chapter he focuses on the fundamental points on which the further theory of “thinking” will be built:



““ ... our memories are consistent in nature. They can be reproduced in the same order as memorized. We cannot get direct access to memories in no particular order. ”

“There are no pictures, videos or sounds in the brain.” All our memories are stored as sequences of patterns (here I don’t even know which word would correctly convey the meaning). Unused memories fade with time. ”

“We can recognize an object even by its part (sight, hearing, touch) and even if the object has undergone significant changes. Our perception has the ability to distinguish characteristics that are not subject to change in the real world. ”

“Our conscious perception actually depends on how we think at the moment. Those. our brain constantly predicts what should happen next, and what we should feel. This expectation affects our perception. ”

“Our memory is not a list of thousands of memorized actions, but rather each of our memorized actions is a carefully designed hierarchy of nested actions. The same hierarchy manifests itself in our ability to recognize objects and situations. ”



Reading this, I had the feeling that I had already seen it somewhere. Well, yes, of course - Jeff Hawkins, On Intelligence. Och advise, by the way. Actually, this is the set of certain requirements brought together, which the inventors of artificial intelligence are trying to embody in their creations.



And so we smoothly approached the third chapter: the neocortex model - the theory of thinking based on pattern recognition. The chapter begins with a rather interesting assessment of the capabilities of the neocortex (cerebral cortex) based on the number of neurons, neural columns and 100'000 “pieces of knowledge”. I do not know where he has such information, but the calculation is approximately as follows: Kasparov learned 100'000 chess positions, Shakespeare composed using 100'000 language constructions, a typical medical specialist knows about 100'000 concepts, and when you consider that many concepts are stored in excess (100 to 1, as he claims), then the human brain can learn 10 million patterns. But by simple statements, he and this number bring up to 300 million. And this coincides with the number of “elementary resolvers”, which he estimates as consisting of about 100 neurons each (30 billion neurons in the neocortex: 500,000 cortical columns with 60,000 neurons each) in each). He gives these estimates without references, so it looks very attractive, but beautiful. Well, it is clear that the elemental recognizer is a perceptron with weights and everything is organized into a hierarchical neural network. And he discusses the whole chapter about the role of patterns and their discriminators, how they are related to each other and influence each other, ending with the mention of his design of a voice recognizer based on hierarchical hidden Markov models: the approach seems to work.

Briefly, the whole chapter can be described as follows: patterns -> elementary resolvers -> neocortex -> Profit.



The fourth chapter has already begun to bring surprises. Describing the biology of the neocortex, he refers to the Blue Brain Project and to the words of its creator, Henry Markram, about groups of neurons found in which synaptic connections and their properties are well predictable and non-random. Moreover, Markram himself assigns them the role of elements of innate memory, and Kurzweil - the role of those “elementary discriminators” (in these groups there are about 100 neurons). Further in this chapter, a 3-dimensional data highway goes through the entire neocortex, opened in March of this year, about neuroplasticity, provided by a single signal processing method in the whole neocortex and about the possible role of dendritic spines in training.



The fifth chapter is fully devoted to the “old” brain, the fact that it is not the neocortex. Ways of sensory information, the role of the thalamus, what is the hippocampus, the role of dopamine and serotonin for fear and pleasure, in general, what we discussed . Interesting, but very concise. As a conclusion to the chapter - for the tasks described above, we do not need to study the old brain at all. Well, he knows better.



Chapter Six is ​​a short chapter on human abilities, creativity, and love. In fact, nothing. Want to learn interesting - read AlexeyR .



And now, to the seventh chapter, we finally arrived at the stated in the title: the digital neocortex from the biological one. Much of the history of computer modeling of the brain, where neural networks have gone, the role of classifiers and hidden Markov models, a little about genetic algorithms and how to use them, a bit about Lisp and Jeff Hawkins. A lot about Watson and Jeopardy !, but not how it works (about it a little bit at all), but how cool and important it is! A whole section is “Brain Creation Strategy”, which summarizes the methods and how to use them and what else it would be nice to add. Very descriptive chapter, practically no specifics.



Chapter 8 - The brain is like a computer. The head of Thuring, von Neumann and Babbidzhe. In general, I was told this at school in computer science. And you?



And then philosophy begins: Chapter 9 - thought experiments with the brain. Here we learn what consciousness is through examples of zombies and qualia , about Penrose with his quantum theory of consciousness (Ray does not approve of him), about faith, the difference in perceptions of the world by Eastern and Western people, free will, determinism / predictability of life, cloning of consciousness. Very interesting topics for a civilized conversation among friends, but the stated topic was about the creation of the brain, and not how this brain will fit into our reality.



Chapter 10 in this book is added solely because Ray loves to make predictions :). It is called so - the law of the exponential growth of technology (loosely translated). And he discusses his old predictions from The singularity is near.



Well and finally chapter 11 - objections, i.e. what other people say about his approach. And most of the chapter is about Paul Allen. Yes, the very one . They have a long-time mezhduosboychik. Paul was wrong on all sides.



Here is such a book. Very artistic. The practical value, especially for those who, for example, took an online course of AI at Stanford, tends to zero, but for general development as a book before bedtime - completely.