« Reading List: Feedback | Main | Reading List: Castigo Cay »

Thursday, February 20, 2014

Reading List: How to Create a Mind

Kurzweil, Ray. How to Create a Mind. New York: Penguin Books, 2012. ISBN 978-0-14-312404-7.
We have heard so much about the exponential growth of computing power available at constant cost that we sometimes overlook the fact that this is just one of a number of exponentially compounding technologies which are changing our world at an ever-accelerating pace. Many of these technologies are interrelated: for example, the availability of very fast computers and large storage has contributed to increasingly making biology and medicine information sciences in the era of genomics and proteomics—the cost of sequencing a human genome, since the completion of the Human Genome Project, has fallen faster than the increase of computer power.

Among these seemingly inexorably rising curves have been the spatial and temporal resolution of the tools we use to image and understand the structure of the brain. So rapid has been the progress that most of the detailed understanding of the brain dates from the last decade, and new discoveries are arriving at such a rate that the author had to make substantial revisions to the manuscript of this book upon several occasions after it was already submitted for publication.

The focus here is primarily upon the neocortex, a part of the brain which exists only in mammals and is identified with “higher level thinking”: learning from experience, logic, planning, and, in humans, language and abstract reasoning. The older brain, which mammals share with other species, is discussed in chapter 5, but in mammals it is difficult to separate entirely from the neocortex, because the latter has “infiltrated” the old brain, wiring itself into its sensory and action components, allowing the neocortex to process information and override responses which are automatic in creatures such as reptiles.

Not long ago, it was thought that the brain was a soup of neurons connected in an intricately tangled manner, whose function could not be understood without comprehending the quadrillion connections in the neocortex alone, each with its own weight to promote or inhibit the firing of a neuron. Now, however, it appears, based upon improved technology for observing the structure and operation of the brain, that the fundamental unit in the brain is not the neuron, but a module of around 100 neurons which acts as a pattern recogniser. The internal structure of these modules seems to be wired up from directions from the genome, but the weights of the interconnections within the module are adjusted as the module is trained based upon the inputs presented to it. The individual pattern recognition modules are wired both to pass information on matches to higher level modules, and predictions back down to lower level recognisers. For example, if you've seen the letters “appl” and the next and final letter of the word is a smudge, you'll have no trouble figuring out what the word is. (I'm not suggesting the brain works literally like this, just using this as an example to illustrate hierarchical pattern recognition.)

Another important discovery is that the architecture of these pattern recogniser modules is pretty much the same regardless of where they appear in the neocortex, or what function they perform. In a normal brain, there are distinct portions of the neocortex associated with functions such as speech, vision, complex motion sequencing, etc., and yet the physical structure of these regions is nearly identical: only the weights of the connections within the modules and the dyamically-adapted wiring among them differs. This explains how patients recovering from brain damage can re-purpose one part of the neocortex to take over (within limits) for the portion lost.

Further, the neocortex is not the rat's nest of random connections we recently thought it to be, but is instead hierarchically structured with a topologically three dimensional “bus” of pre-wired interconnections which can be used to make long-distance links between regions.

Now, where this begins to get very interesting is when we contemplate building machines with the capabilities of the human brain. While emulating something at the level of neurons might seem impossibly daunting, if you instead assume the building block of the neocortex is on the order of 300 million more or less identical pattern recognisers wired together at a high level in a regular hierarchical manner, this is something we might be able to think about doing, especially since the brain works almost entirely in parallel, and one thing we've gotten really good at in the last half century is making lots and lots of tiny identical things. The implication of this is that as we continue to delve deeper into the structure of the brain and computing power continues to grow exponentially, there will come a point in the foreseeable future where emulating an entire human neocortex becomes feasible. This will permit building a machine with human-level intelligence without translating the mechanisms of the brain into those comparable to conventional computer programming. The author predicts “this will first take place in 2029 and become routine in the 2030s.”

Assuming the present exponential growth curves continue (and I see no technological reason to believe they will not), the 2020s are going to be a very interesting decade. Just as few people imagined five years ago that self-driving cars were possible, while today most major auto manufacturers have projects underway to bring them to market in the near future, in the 2020s we will see the emergence of computational power which is sufficient to “brute force” many problems which were previously considered intractable. Just as search engines and free encyclopedias have augmented our biological minds, allowing us to answer questions which, a decade ago, would have taken days in the library if we even bothered at all, the 300 million pattern recognisers in our biological brains are on the threshold of having access to billions more in the cloud, trained by interactions with billions of humans and, perhaps eventually, many more artificial intelligences. I am not talking here about implanting direct data links into the brain or uploading human brains to other computational substrates although both of these may happen in time. Instead, imagine just being able to ask a question in natural language and get an answer to it based upon a deep understanding of all of human knowledge. If you think this is crazy, reflect upon how exponential growth works or imagine travelling back in time and giving a demo of Google or Wolfram Alpha to yourself in 1990.

Ray Kurzweil, after pioneering inventions in music synthesis, optical character recognition, text to speech conversion, and speech recognition, is now a director of engineering at Google.

In the Kindle edition, the index cites page numbers in the print edition to which the reader can turn since the electronic edition includes real page numbers. Index items are not, however, directly linked to the text cited.

Posted at February 20, 2014 23:16