Ray Kurzweil is a futurist who may be most famous to many people for his book “The Singularity is Near”, where he generated some controversy because he said we are soon reaching a point (the singularity) where humans can, in theory, live forever.
I have never read any of his books, because I have never been too interested in what futurists say, as in general they have been wrong. Being a “futurist” is kind of like being a fortune teller. You might get lucky, but in general you are just guessing.
However, I decided to give this book a try because of my day job. Mr. Kurzweil has created technologies which have essentially morphed into what is in Siri and Google Voice Search – speech recognition, and in my day job I will be working on this.
The central thesis of this book is Mr. Kurzweil’s assertion that the neocortex of a human brain is essentially a set of pattern recognizers which are repeating and work at different levels. A pattern recognizer can be a simple as, say, recognizing the cross-bar of the capital letter “A” and saying “this could be a letter A”. Another recognizer could recognize the left bar, and a third the right bar, and when they fire together, a recognizer at a higher level will say “ah, this must be an “A”. And this goes on and on to create recognizers for words, then phrases, then concepts.
He calls this the “Pattern Recognition Theory of Mind” (PRTM). To back up his case, he points to many recent discoveries in the medical field that show that the neocortex is not this random rats nest of neurons generating random axons to random dendrites, (axon being the messenger piece, and dendrite being the receiver piece), but rather a regular, highly repeating structure where of neurons in clusters, with connections between the clusters.
Mr. Kurzweil submits that these clusters are pattern recognizers, who learn a specific thing (such as the crossbar of a capital A) and that groups of these work together to create a letter A in your mind. In addition, these pattern recognizers work downward. As an example, for the word “CAT”, the brain sees a letter C at a higher level and sends a signal down to the pattern recognizers for A saying “I’m kind of expecting an A here, such that even though the A may not be written very clearly, it still fires. This is how we can see, say, a picture of Einstein that is half covered, or is a cartoon drawing of Einstein, but we still recognize it as Einstein.
To further back up this claim, Kurzweil goes into some explicit detail on how he developed speech recognition algorithms, by using HHMMs (Hierarchical Hidden Markov Models), which is a very strong concept used in modern Artificial Intelligence applications. Watson, for example, the computer that beat contestants at jeopardy, and before that Big Blue, which beat a world class chess champion at chess, used HHMMs as part of their core programming. Watson, for example, was not programmed with all the world’s information – it was given basic rules of grammar and structure, and then was “set loose” to read every article on Wikipedia on its own. Based solely on this, Watson was able to figure out puns, for example – they weren’t hand programmed in.
Kurzweil then spends some time talking about HHMMs and a concept called “GA” (genetic algorithms). The way to make his speech recognition program work better was to build a set of HHMMs, have them work on recognizing speech, and then taking the best results from the different sets and, well, “mating” them – taking some of the algorithm from model A and some of the algorithm from model B. Do this to create new sets, and drop all the ones that “failed”. Run the recognition again, and repeat. Eventually you get something that is much better than when you have a bunch of linguists program a speech recognizer by hand. The resulting algorithm built itself – made itself smarter, and made itself smarter than humans.
Which then begs the question – why does Siri, and Google Voice Search, get things wrong? This is a question not of it being “dumb” or even being “not conscious”… it is a question of resources. Neither Siri nor Google have applied the richest set, as this requires too much compute power. But as Kurzweil shows, this is just a matter of time – technology progresses at an exponential rate.
At this point Kurzweil takes an aside to talk about consciousness and free will. What are they? And would a computer program ever develop them? He spends some time knocking down arguments for why a computer could never be conscious. One of the main arguments made by people is that the brain is doing “quantum computing” in its axons, and since quantum computing is “many states at once”, there is no way a computer that is operating on “1” and “0” states can duplicate it. He shows that this neurons doing quantum math is, really, nonsense, and that even if it weren’t, that just means it will take longer for a computer to get there (computers are edging into the realm of quantum computing.
But then he does a great explanation of consciousness as discussed by philosophers. Any discussion of consciousness or free will ends up using definitions that are circular. For example, free will is defined as the freedom to make a decision without constraints – but since “freedom” is part of the definition, that’s a circular argument. This gets a little long winded for me, because I have never cared too much for philosophy, but it does serve the point he is trying to make. Given that we can’t really define what consciousness is, it is pretty difficult to say whether a computer is or is not conscious.
In general, pretty good book. While he doesn’t use a lot of math in the book, there is enough math in it that it can get frustrating if you don’t care for math. While I like math, even I got a little bored in these parts. But overall, it made a really great impression on me about how the brain works, and how it should therefore not be too difficult to model a human brain and thus, consciousness. As he points out, the limits to our intelligence are how much neocortex we have – how many of the pattern recognizers we can fit into our skull. Our brain is totally dominated by the neocortex, but there is only so much we can have. Given that, our technology – our ability to create services in the cloud, allows us to extend our consciousness beyond our forehead. Eventually these technologies will be so small as to be ubiquitous. We already carry an incredibly smart agent in our pocket (the cell phone) that can understand (with limitation) what we say to it. Things like Google Now are using even more information about us (through our use of Google services such as Gmail, and Google Search), to give you information on things you care about without you even asking.
So, if you can get by with some of the math without losing your own mind, I highly recommend the read.