Friday, July 31, 2015

Ray Kurzweil's How To Create a Mind

I've been reading Ray Kurzweil's newest book from 2012.  I know I am not a distinguished reviewer, but I've worked in the same field and I have some overlapping experience.  

I don't write extensive reviews because in the end it is down to the reader to judge and you are the one with the least experience on the subject matter. You will either take my word for it or you won't. If I am to make an impression I will do so by making a few short sharp points. To make a longer point I may as well write my own book.

I will make a three point complaint.

The first point is that the title of the book demonstrates he did not deliver an honest and complete description of a mind.  His title contains three disingenuities at the same time. He does not describe how to make a complete mind. He does not describe the one secret of human thought. He does not know the secret of a human mind, his title seems to be a publishing exaggeration for sales.  In the book once you've read it you would notice he claims to predict when a human mind - ready computer system will act consciously. He is basing it on extrapolating his known computing elements. He is estimating we will be at a complete mind when we reach enough pattern recognizers. I will go into my complaints in greater detail below.  Suffice to say that his claims are not proven.

Here is the simple way to tell if he knows how to create a mind. If he knows a valuable secret like artificial intelligence, then why would he sell it to everyone to copy for $17? Why would he need to write a book about it?  He does know a great deal but not the key pieces. No one does.

The exaggerations made in this regurgitated book are an embarrassment for any serious academic.  He has forgotten his roots, and has not done a credible job of completing the design for human thought as he claims.  He estimates but that like any other supposition, extrapolation, and so on is just a guess. He believes his educated guess is worth more, but he does not underline where his assumptions might be wrong. He is confident in his guesses. But guesses are not sufficient, and his design is not proven.

The second point is that his main theory does not deliver a complete human mind.  His approach, that a human brain is predominantly a pattern recognition system in the neocortex, and that by putting enough neocortex together at the right processing speed will make a human mind is just as misguided as many other incomplete theories.

His approach is equivalent to making a car out of all the parts we understand, like getting a box full of  40,000 parts and you know about 30,000 of them, and you assemble them like a copied car sitting in the drive way and expect it will drive.  Now suppose that you didn't get tires in your box of parts so you substitute another kind of tire made from glass.  You install the glass tires and try to drive. They look the same and they roll the same so they are close enough, right?

Even if you got the engine to work without understanding physics, when you drive and it can't turn or accelerate or brake because you don't understand friction have you really duplicated a car?  That is the hallmark of a scientific model that does not account for understanding of all the important principles involved in a complex model. When someone makes a model that is an approximation it works well over a smaller region of the original, sometimes it won't work at all and it fails in unexplained ways.  An approximation is not equivalent, it is approximately equivalent.

My third complaint is that he discards the reality that we don't know the artificial intelligence principle, like the Bernoulli effect for aviation. He claims that since we know how the neocortex works with lego blocks of pattern recognizers then it's only a matter of collecting enough of them to make a mind. Even his approach is close to my car from a box of part analogy. But what if the missing pieces are the critical ones?  It is my supposition is that we don't have the final missing piece. Not the missing link, but the missing principle that explains all of brain functions and also the abstracted brain and its ways to learn, extrapolate, be creative, and so on. 

Without the missing piece, it might be forever before we can replicate a human mind.

In sum, while he presents some interesting new information along with some regurgitated elements of older books, and some truly disturbing parts when he is either product promoting ad nauseum for his speech recognition work, or he is claiming cyborg future vision in the final chapter, he does not deliver what he claims. This book is not that useful despite all the reputation he holds. If you want a better book that explains a less ambitious portion of the brain, then read David Marr's Vision as his book does cover the entire abstracted organization and purpose of the human vision system. It does not claim to know all the vision system but it does provide you with a solution. It does not claim more than it delivers and it will outlast the final understanding of the seeing brain.