How to Create a Mind: The Secret of Human Thought Revealedby Published 13 Nov 2012
|How to Create a Mind: The Secret of Human Thought Revealed.pdf|
The bold futurist and bestselling author explores the limitless potential of reverse-engineering the human brain
Ray Kurzweil is arguably today’s most influential—and often controversial—futurist. In How to Create a Mind, Kurzweil presents a provocative exploration of the most important project in human-machine civilization—reverse engineering the brain to understand precisely how it works and using that knowledge to create even more intelligent machines.
Kurzweil discusses how the brain functions, how the mind emerges from the brain, and the implications of vastly increasing the powers of our intelligence in addressing the world’s problems. He thoughtfully examines emotional and moral intelligence and the origins of consciousness and envisions the radical possibilities of our merging with the intelligent technology we are creating.
Certain to be one of the most widely discussed and debated science books of the year, How to Create a Mind is sure to take its place alongside Kurzweil’s previous classics which include Fantastic Voyage: Live Long Enough to Live Forever and The Age of Spiritual Machines.
"How to Create a Mind: The Secret of Human Thought Revealed" Reviews
I'm just going to warn everyone at the offset: this book triggered my grumpy, cane-waving, "you kids get off my lawn" reflexes pretty hardcore. So, buckle up.
If you ever need a really clear example of how intelligence and wisdom are not the same thing, this book is a great place to get started. I don't for an instant doubt that Ray Kurzweil is a very, very smart guy. (Almost certainly smarter than I am.) The problem is that, like quite a lot of people who have had a super-abundance of success--a dearth of healthy failure leads to a superabundance of confidence. Basically everything that is wrong in this book stems from that. So let's get started.
Two of Kurzweil's motivating points are that:
1. we don't really need to understand principles to use them
2. more complicated things are sometimes simpler than less complicated things
For the first, he cites the fact that while an exact scientific explanation of Bernoulli's principle is still controversial (at best), we have airplanes. So, the theory goes, if his Pattern Recognition Theory of Mind (PRTM) works then we can build AI without fully grasping it just like we build airplanes while still pondering Bernoulli's principle. Except, of course, that if you went back in time 2,000 years and explained Bernoulli's principle to people, they couldn't exactly go out and build a 747. Same principles here: even if PRTM is correct, the idea that it's sufficient to build human-level AI is totally unsubstantiated. (We'll get to some specific deficits shortly.)
As for the second, here's his argument:
Let's think about what it means to be complex. We might ask, is a forest complex? The answer depends on the perspective you choose to take. You could note that there are many thousands of trees in the forest and that each one is different. You could then go on to note that each tress has thousands of branches and that each branch is completely different. Then you could proceed to describe the convoluted vagaries of a single branch. Your conclusion might be that the forest was complexity beyond our wildest imagination.
But such an approach would literally be a failure to see the forest for the trees. Certainly there is a great deal of fractal variation among trees and branches, but to correctly understand the principles of a forest you would do better to start by identifying the distinct patterns of redundancy with stochastic (that is, random) variables that are found there. It would be fair to say that the concept of a forest is simpler than the concept of a tree.
This is analogy is, shall we say, strained. If all you care about is geometry, then yes: you could use procedural generation to generate a reasonable facsimile of a forest without trying to minutely recreate an actual forest. (This is the idea behind the video game No Man's Sky, which uses procedural generation to create a universe "which includes over 18 quintillion (1.8×1019) planets, many with their own sets of flora and fauna."
However, (1) that's how to create a forest in general, not any particular forest, and (2) what's the legitimate excuse for reducing a tree or a forest to merely their geometric shapes? A tree is an organism, and a forest is a superorganism of even greater complexity (even if you only consider the relationships between the various trees and ignore all the other creatures inhabiting it.)
In short, this is just wishful thinking presented as an argument.
There's an awful lot of pseudoscience like this in the book. He spends some time estimating the total number of patterns that a human brain needs to memorize, patterns for everything from the shape of the letter "a" to rules for driving safely. Unsurprisingly, his estimate of the number of patterns we need to memorize and his estimate for the number of discrete pattern-recognition unites in the neocortex coincide. This is convenient for his theory, but useless for any other purpose because he had defined his terms so loosely (if at all) that the explanation is entirely a black-box, while the method used to derive it is no more scientific than the infamous Drake equation.
Another argument he repeats several times is that the brain really can't be that complex because there's only but so much information related to the brain in our DNA. This is an extremely problematic assertion, because it leaves out the very serious possibility of pointers.
Think about it this way: we can easily compute how much memory it takes to send a string of English characters. If you use a typical encoding, you spend about 1-2 bytes per character, so a thousand characters is about a kilobyte. Now consider that I encode random gibberish and send it. How much information have I sent? Or imagine I write something meaningful, but that I send it to someone who doesn't speak English? How much information has been transmitted? Now imagine that I send it to someone who does speak English. Obviously this person--receiving a meaningful message--gets a lot more information out of what I've sent than the previous two. And so obviously measuring how much information is available for transmission isn't the same thing as measuring how much information is received. The English speaker is interpreting my message against a vast library of linguistic data that they already have in mind, and so they're getting a lot more out of it. Or, to make this example really extreme, suppose that my message says, "You should look up this article on wikipedia" and then provides a URL.
This is what I really mean by a reference. It's possible to send a small message that refers to a much larger amount of information stored elsewhere. Effectively, this is what DNA is doing, since it's basically saying "here's how to build proteins" and then relying on the information encoded in physics--which dictates the behaviors and interactions of those proteins--to reference a vast library of information. How much data it takes to send instructions to build a brain via DNA and how much data we would need to replicate a mind in some other substrate are entirely different questions. The only way we could be assured of needing no more information than is available in the DNA is if we were actually building a biological brain or, at least, simulating one at the atomic level, which is exactly what Kurzweil insists we don't have to do. In other words: more wishful thinking.
The real irony here, of course, is that Kurzweil refers to these concepts in the book. He understands that human intelligence is important because it allows us to store information "in the cloud." Historically, this meant social knowledge and culture. Instead of transmitting knowledge via DNA, we transmitted it via spoken language, and so we could store a lot more. Than we figured out writing, etc. He's familiar with all of this, so he should understand that the same could be true of DNA.
But the most egregious problems occur when Kurzweil ventures outside of science entirely and starts talking philosophy. (This is a common problem with science writers, by the way. They really tend to lack a fundamental sense of humility when treading outside their own specialized domains.) For example, he has a chapter on theories of consciousness where--after summarizing some of the competing views--he basically waves the entire topic aside: "These theories are all leaps of faith... where consciousness is concerned 'ya gotta have faith'" It's odd for someone to dismiss so glibly the entire topic of consciousness and yet still take such a strong stance that we'll have conscious machines by 2029.
Perhaps the most embarrassing section was his discussion of Descartes' famous "cogito ergo sum." Kurzweil writes:
is famous "I think, therefore I am" is generally interpreted to extol rational thought, in the sense that, "I think--that is I can perform logical thought--therefore I am worthwhile"
This is an absurd misinterpretation of Descartes' statement that I have literally heard nowhere. I can imagine a remarkably ignorant person who decided to invent meaning for a quote without bothering to look it up online might come up with such a silly theory, but to claim it is "generally interpreted" this way is only to reveal a deep chasm of ignorance on fundamentals of Western philosophy. To his credit, Kurzweil goes on to say that "reading this statement in context of his other writings I get a different impression," whereupon he gives the correct summary of Descartes' point. However, getting the point correct doesn't really make up for claiming as your own interpretation the meaning of a phrase that (1) is abundantly clear in context and (2) appears in introductory philosophy textbooks everywhere.
As I stated: it's probably just a case of overconfidence. Kurzweil is probably really quite brilliant in his area of expertise (e.g. hierarchical hidden Markov modes, but it certainly erodes his credibility when he speaks so confidently and incorrectly on matters that are really pretty basic within their own domain.
This is compounded when, a few pages later, he seriously attempts to defend his Law of Accelerating Returns (a generalized version of Moore's Law) as being just as much a real "law" as the laws of thermodynamics. I'll leave out a detailed rebuttal of this point, because I think for most people the silliness--and vanity--of this position are self-evident.
All of this culminates in his final attack on John Searle and his Chinese room thought experiment. Searle is one of the most respected philosophers alive today, and Kurzweil is someone who thinks stating the obvious and accepted interpretation of philosophy's most famous three words is somehow his own invention. Naturally, this does not end well.
Searle's Chinese room thought experiment is based on a simple setup: imagine a room with a library full of Chinese symbols listed in a table of input / output. Inside this room sits a man who speaks not a single word of Chinese. Occasionally, someone will write a question (in Chinese) and submit it through a slot. The man then looks through his library of symbols to find an input that matches. When he does, he copies down the output and slides that paper back out of the window.
To an outside observer, it looks as though they can simply ask the room a question in Chinese and get an answer. So they might naturally assume that the person in the room speaks fluent Chinese. Of course, having to wait minutes or even hours to get a reply might spoil the illusion, but the point of this example is to make an analogy for a computer, so you can imagine the person looking up the Chinese symbols is really, really fast.
Of course, we know that he doesn't understand a word. And so--according to Searle--the fact that an AI program could (for example) pass the Turing test wouldn't guarantee that the AI understood anything at all.
Kurzweil's rebuttal to the most famous argument against strong-AI is to simply state that--while the man doesn't understand Chinese--the man and the room taken together do. This is not bad as a starting point to try and respond to Searle, but it's just that: a starting point. It's not clear to me at all that a library of books "understands" information just because it contains information. And it's certainly not clear to me what it means to say that a library + a librarian = a unified, holistic intelligence. (Especially if the librarian happens to be unable to read any of the words in his library!)
Suffice it to say: I didn't find this book very convincing.
I find the overall topic interesting, and I do think that Kurzweil's explanation of how the neocortex works is entirely plausible. I found it convincing, anyway. This PRTM (pattern recognition theory of mind) seems entirely plausible as an explanation for how the neocortex handles knowledge--both learning and recall--and I do think that his basic aspiration (to use that as a basis for constructing cybernetic brain-enhancements) has some realistic promise. I say, some because I think he's skipping over some really, really hard problems, especially how to integrate an artificial neocortex with a biological one. The neocortex is impressively plastic, but not infinitely so, and it's not at all clear to me that this is a trivial problem.
Moreover, creating an auxilliary neo-cortex is very, very, very far from creating a stand-alone AI. As Kurzweil admits, even if you constructed an artificial brain (he seems to think you could get away with just a neocortex) you're going to have to teach it. But he views this as a symbol matter of pattern recognition. This is obviously false, because a lot of what is required for growing a healthy human mind includes things like love and empathy. He suggests that waiting to teach a nascent AI in real time would be tiresome, so we should speed up the clock cycle and fast-forward our artificial brain through 10-20 years of development in 1-2 years of real time. Would anyone like to propose a method of providing a simulated version of 10-20 years of love and healthy attachment to a computer simulation at 10x speed? Because we know--from tragic natural experiments--that a lack of love and nurture leads to severe developmental problems both emotionally and intellectually.
There's some really, really interesting material here. And if Kurzweil was willing to show some humility in dealing with experts outside his field--and maybe a little bit of humility with his own visions--it could have been a fascinating and influential book. As it stands, however, his achievements are overshadowed by his unjustified arrogance.
I like Kurzweil. But I thought he did a little too much boasting and did not provide enough details.
First half of the book: it appears that we can model the brain with hierarchical hidden Markov models better than we can with neural nets. Some back of the envelope calculations show that Hidden Markov models may contribute to the functioning of the brain. Ok, so far so good.
Second half of the book: wildly uneven coverage of a wide range of topics in neuroscience philosophy, such as identity, free will, and consciousness.
Kurzweil likes to frequently mention all of the contributions that he has made to AI. I think this could have been toned down a little bit. Back in year XXX, I was one of the first to do XYZ.
He has some good ideas in the first part, but I don’t think he comes close to explaining how to create a mind.
*A full executive summary of this book is available here: http://newbooksinbrief.com/2012/11/27...
When IBM's Deep Blue defeated humanity's greatest chess player Garry Kasparov in 1997 it marked a major turning point in the progress of artificial intelligence (AI). A still more impressive turning point in AI was achieved in 2011 when another creation of IBM named Watson defeated Jeopardy! phenoms Ken Jennings and Brad Rutter at their own game. As time marches on and technology advances we can easily envision still more impressive feats coming out of AI. And yet when it comes to the prospect of a computer ever actually matching human intelligence in all of its complexity and intricacy, we may find ourselves skeptical that this could ever be fully achieved. There seems to be a fundamental difference between the way a human mind works and the way even the most sophisticated machine works--a qualitative difference that could never be breached. Famous inventor and futurist Ray Kurzweil begs to differ.
To begin with--despite the richness and complexity of human thought--Kurzweil argues that the underlying principles and neuro-networks that are responsible for higher-order thinking are actually relatively simple, and in fact fully replicable. Indeed, for Kurzweil, our most sophisticated AI machines are already beginning to employ the same principles and are mimicking the same neuro-structures that are present in the human brain.
Beginning with the brain, Kurzweil argues that recent advances in neuroscience indicate that the neocortex (whence our higher-level thinking comes) operates according to a sophisticated (though relatively straightforward) pattern recognition scheme. This pattern recognition scheme is hierarchical in nature, such that lower-level patterns representing discrete bits of input (coming in from the surrounding environment) combine to trigger higher-level patterns that represent more general categories that are more abstract in nature. The hierarchical structure is innate, but the specific categories and meta-categories are filled in by way of learning. Also, the direction of information travel is not only from the bottom up, but also from the top down, such that the activation of higher-order patterns can trigger lower-order ones, and there is feedback between the varying levels. (The theory that sees the brain operating in this way is referred to as the Pattern Recognition Theory of the Mind or PRTM).
As Kurzweil points out, this pattern recognition scheme is actually remarkably similar to the technology that our most sophisticated AI machines are already using. Indeed, not only are these machines designed to process information in a hierarchical way (just as our brain is), but machines such as Watson (and even Siri, the voice recognition software available on the iPhone), are structured in such a way that they are capable of learning from the environment. For example, Watson was able to modify its software based on the information it gathered from reading the entire Wikipedia file. (The technology that these machines are using is known as the hierarchical hidden Markov model or HHMM, and Kurzweil was himself a part of developing this technology in the 1980's and 1990's.)
Given that our AI machines are now running according to the same principles as our brains, and given the exponential rate at which all information-based technologies advance, Kurzweil predicts a time when computers will in fact be capable of matching human thought--right down to having such features as consciousness, identity and free will (Kurzweil's specific prediction here is that this will occur by the year 2029).
What's more, because computer technology does not have some of the limitations inherent in biological systems, Kurzweil predicts a time when computers will even vastly outstrip human capabilities. Of course, since we use our tools as a natural extension of ourselves (figuratively, but sometimes also literally), this will also be a time when our own capabilities will vastly outstrip our capabilities of today. Ultimately, Kurzweil thinks, we will simply use the markedly superior computer technology to replace our outdated neurochemistry (as we now replace a limb with a prosthetic), and thus fully merge with our machines (a state that Kurzweil refers to as the singularity). This is the argument that Kurzweil makes in his new book 'How to Create a Mind: The Secret of Human Thought Revealed'.
Kurzweil lays out his arguments very clearly, and he does have a knack for explaining some very difficult concepts in a very simple way. My only objection to the book is that there is a fair bit of repetition, and some of the philosophical arguments (on such things as consciousness, identity and free will) drag on longer than need be. All in all there is much of interest to be learned both about artificial intelligence and neuroscience. A full executive summary of this book is available here: http://newbooksinbrief.com/2012/11/27...
A podcast discussion of the book will be available soon.
Beyond some spurious dialog of computer modeling, the book is cleanly written and well-argued. The chapter on consciousness offers an amazing discussion of how a computer can (or can’t) replicate a human mind. The author finishes by taking on objections to his ideas. Highly recommended.
While the brain has been considered by many to be beyond the scope of comprehension, history is replete with claims of what couldn’t be done. How to Create a Mind offers a thoroughly supported argument for the eventual reverse engineering of the human brain.
Kurzweil's book offers an overview of the biological brain and briefly overviews some attempts toward replicating its structure or function inside the computer. He also offers his own high-level ideas that are mostly a restatement of what can already be found in other books (such as Hawkins' On Intelligence) with a few modifications (he admits this himself though at one point, for which he gets bonus points). Finally, he applies his Law Of Accelerating Returns (LOAR) to field of AI and produces some predictions for the future of this field.
The good: Nice thought experiments section, nice overview of the biological brain (both old brain/cortex and their function), reasonably ok philosophical mambo jambo parts about consciousness and whether it is possible for a computer to be a mind (if you're into that), some analysis of relevant computational trends. By the end, you're almost convinced we're almost there!
The bad: First, his own theories are extremely vague and half-baked (though I forgive this. If he knew more he would be busier with things other than writing this book) and essentially reduce to some form of Hierarchical Hidden Markov Model. That's not especially exciting, I think most researchers in the field will agree on such high-level things. I also find it puzzling that he claims to be talking about the mind in its entirety, but then his exposition focuses almost entirely on temporal modeling/prediction aspects and mostly ignores a lot of other magical components of a mind, such as a flexible and efficient knowledge representation / inference engine, or a reinforcement learning - like actor /critic system that surely exists somewhere at the core of all of our learning and reasoning.
All in all, I would recommend this book to anyone who's interested in some pointers to our efforts to replicate a brain in the computer, who wants to learn a bit about the biological brain, or who's into the philosophy of it all.