Aug 23 2010

Kurzweil vs Myer on Brain Complexity

There is an interesting blog debate going on between PZ Myers and Ray Kurzweil about the complexity of the brain – a topic that I too blog about and so I thought I would offer my thoughts. The “debate” started with a talk by Kurzweil at the Singularity Summit, a press summary of which prompted this response from PZ Myers. Kurzweil then responded here, and Myers responded to his response here.

Futurism

You can read the exchange for all the details. I want to focus on just a couple of points – predicting our efforts to reverse engineer the brain, and the question of how complex the brain is.  Kurzweil has predicted in the past that we will reverse engineer the brain – model its function in a computer, basically – by 2030. It was reported that in his talk he said 2020, but Kurzweil has clarified that this is not correct, he said 2030, sticking to his earlier predictions.

That’s a minor (but interesting) point, and Myers points out that it was not the focus of his original criticism. I agree with Kurzweil on some basic principles. First, we do have an active research program that is using computer modeling to reverse engineer the brain. These efforts are progressing nicely, and I do think that eventually they will succeed. I also agree that some technologies progress at an exponential rate, and they surprise those who were making predictions based upon a linear progression. Kurzweil gives an excellent example of this – the genome project. This project started out very slow, and many thought it was lagging behind predictions, but as technology improved the effort to decode the human genome accelerated geometrically and actually finished years ahead of schedule. Now we can decode the genome of other species in a fraction of the time, and the pace continues to accelerate.

So Kurzweil has a legitimate point here – information-based technologies are accelerating, and if you account for that acceleration you get a better handle on predicting its future course. I do think, however, that Kurzweil is cherry-picking a bit also, for some information-based technologies have fallen short of prediction, such as speech recognition (an area of his particular expertise). Recognizing human speech works, but the technology has seen diminishing, rather than accelerating, returns in terms of accuracy, and this has delayed its adoption – which is not nearly as much as Kurzweil predicted in the past.

I think the example of speech recognition represents a factor that Kurzweil, in my opinion, seems to underappreciate. While our information tools may get better at an accelerating rate, some problems become exponentially more difficult as you try to eke out incremental gains. In other words, it seems that for some technologies (to use symbolic figures) each 1% improvement is 10 times more difficult than the previous incremental improvement. This offsets our exponential progress. The complexity of the genome project was linear – decoding that last 10% was as difficult as the first 10%, so it was the perfect example for Kurzweil. But other problems, like understanding how the brain works, are not linear in complexity. As our knowledge of the brain deepens, we are getting to greater and greater levels of complexity.

Further, while I think Kurzweil’s characterization of technological progress is generally correct when you consider the broad brushstrokes of advancement, it is very difficult to apply them to any individual technology. There are hurdles, roadblocks, and breakthroughs with any individual technology or scientific problem that are impossible to predict.

On the point of predicting the future I am somewhat between Myers and Kurzweil. Kurzweil has some legitimate points to make, but I think he over applies them and cherry picks favorable examples. Myers also has some legitimate criticisms – Kurzweil does not quantify some problems (like how much of the brain we currently understand), and does not account for the fact that we do not know how much we do not know. There may be hidden layers of complexity of brain function we haven’t tapped into yet. But I think that Myers overall is a bit harsh on Kurzweil and does not give partial credit where it is due.

Will we reverse engineer the brain by 2030? I guess we will have to wait and see. Kurzweil gives himself a bit of an out by saying that we will reverse engineer the “basic functions” of the brain – this is vague enough that you can declare victory at any point along the way. You might argue we understand the brain’s basic functions now. I think we will succeed eventually, even to the point of being able to make an artificial brain, but I would not hazard a guess as to when.

Brain Complexity

The more interesting point of contention, and a real teaching point, is the question of how much we can infer about the complexity of the brain by looking at the genome? A separate question is whether or not you can reverse engineer the brain by examining the genome. Here both Myers and Kurzweil agree – you cannot. But Kurzweil says he never made that claim – it was misreported or misinterpreted. So we can put that aside – no one is arguing that the design of the brain is in the genome. You have to examine the brain to reverse engineer the brain.

But Kurzweil is still claiming that we can infer something about how much complexity is in the brain from the genome. He writes:

The amount of information in the genome (after lossless compression, which is feasible because of the massive redundancy in the genome) is about 50 million bytes (down from 800 million bytes in the uncompressed genome). It is true that the information in the genome goes through a complex route to create a brain, but the information in the genome constrains the amount of information in the brain prior to the brain’s interaction with its environment.

This is profoundly problematic, and reflects the fact that Kurzweil truly does not understand the process by which the brain develops. From a developmental point of view – there is no such thing as the brain prior to its interaction with the environment. First – is Kurzweil talking about a newborn infant’s brain? Does he understand the significant differences between that brain and a fully developed adult brain?

I think, to be generous, Kurzweil is trying to differentiate the design of the brain from the information contained within it (our memories, etc.). This could be analogous to a computer vs the software, or reverse engineering a generic human brain vs duplicating PZ Myers’ brain.

But that was never the point at all – the point Myers was making (which I also discussed this week on the SGU) is that the design of the brain is dependent upon interaction with the environment. Myers focused on brain proteins interacting with each other in a complex way, while I focused on the neurological functions of the brain.  The genome provides a set of processes by which brain design unfolds – but that program is dependent upon input from the brain’s environment, which includes the body of which it is part. The basic systems within the brain develop and organize themselves in response to sensory input or use. Our visual cortex requires visual stimulation, binary vision requires seeing with both eyes, our motor system requires use against gravity, our language cortex requires exposure to language, etc.

The process of brain design being a combination of genetic rules laying out neurons and connections in a pattern that is dependent upon feedback from some kind of input adds complexity and information to the brain. So again – what is Kurzweil talking about when he refers to a brain prior to interaction with the environment? He seems not to understand the process of brain development, and therefore he overestimates the degree to which information in the genome constrains information in the brain – or he underestimates the increase in information that derives from this interactive development process. Therefore his basic premise – the brain is not so complex because the genome does not contain that much information – is flawed and invalid (which was Myers original criticism).

Kurzweil adds another line of reasoning to his argument, writing:

For example, the cerebellum (which has been modeled, simulated and tested) — the region responsible for part of our skill formation, like catching a fly ball — contains a module of four types of neurons. That module is repeated about ten billion times. The cortex, a region that only mammals have and that is responsible for our ability to think symbolically and in hierarchies of ideas, also has massive redundancy. It has a basic pattern-recognition module that is considerably more complex than the repeated module in the cerebellum, but that cortex module is repeated about a billion times. There is also information in the interconnections, but there is massive redundancy in the connection pattern as well.

Here again, Kurzweil is grossly underestimating the complexity of the brain based upon some faulty assumptions. I agree with his point that there are modules or patterns in the brain that are repeated billions of times. But they are not simply repeated. You cannot describe this aspect of brain design by simply describing one module then say – repeat 1 billion times. With each repetition there is a novel and meaningful pattern of interconnectedness to other brain regions and to the body. Kurzweil seems to recognize this when he says: “There is also information in the interconnections, but there is massive redundancy in the connection pattern as well.” But he seems to be brushing it off too easily. We cannot assume that the pattern of interconnectedness is a simply redundant pattern.

We also have to consider that the added levels of complexity from the pattern of interconnectedness likely vary from brain region to region. Kurzweil might have a point if you are talking only about the primary visual cortex, for example – where there is a literal grid of neurons that correspond to the visual fields. Here the patterns are somewhat simple and repeated, and it is therefore not surprising that our efforts to reverse engineer these brain regions have progressed the most. But this is the lowest hanging fruit, and should not be considered representative of other brain regions and functions.

If we move to brain regions that subsume our most complex abstract thought and planning, there is no simple somatotopic pattern of neurons whose function we can easily infer. We have no idea, for example, how a pattern of neuronal connections equals a specific word, and connects to our knowledge of how to say the word, how to spell it, what the word means in all its complexity, memories of the word’s use, and its relation to other words and parts of words. But most importantly – we really don’t know yet how complex this problem even is, and so predicting how long it will take to solve the problem strikes me as utter folly.

Conclusion

I find the entire discussion between Myers and Kurzweil to be a fascinating topic, and an opportunity to explore various aspects of neurology in the context of a specific and interesting application – reverse engineering the brain. This amounts to an elaborate thought experiment, but those are a fun way to challenge our understanding of a topic.

Ultimately I come down closer to Myers’ position – Kurzweil does not seem to understand the brain or brain development, at least in certain key aspects, and this dooms his arguments to failure. He would do well to take the criticisms going his way seriously, and also to check his ideas with some actual neuroscientists. Myers, on the other hand, came off too harsh, but that seems to be his style. Kurzweil is an interesting mix of provocative ideas, some interesting insights, but also some serious flaws that border on crankery. This makes him a very intriguing character that I would not casually dismiss, but also I would take everything he says with a skeptical grain of salt.

26 responses so far