Here’s a paper that I just read that makes a very interesting point. The paper by Eric Jonas and Konrad Kording (J&K) has the provocative title “Could a neuroscientist understand a microprocessor?” It tests the techniques of neuroscience by applying them to a structure that we completely understand and asks if these techniques allow us to uncover what we know to be the correct answer. The “model system” investigated are vintage processors used to power video games in very early Apple/Atari/Commodore devices and the question asked is whether the techniques in cog-neuro can deliver an undergrad level understanding of how the circuit works. You can guess the answer: Nope! Here’s how J&K put it:
Here we will try to understand a known artificial system, a historic processor by applying data analysis methods from neuroscience. We want to see what kind of an understanding would emerge from using a broad range of currently popular data analysis methods. To do so, we will analyze the connections on the chip, the effects of destroying individual transistors, tuning curves, the joint statistics across transistors, local activities, estimated connections, and whole brain recordings. For each of these we will use standard techniques that are popular in the field of neuroscience. We find that many measures are surprisingly similar between the brain and the processor and also, that our results do not lead to a meaningful understanding of the processor. The analysis cannot produce the hierarchical understanding of information processing that most students of electrical engineering obtain. We argue that the analysis of this simple system implies that we should be far more humble at interpreting results from neural data analysis. It also suggests that the availability of unlimited data, as we have for the processor, is in no way sufficient to allow a real understanding of the brain. (1)
This negative result should, as J&K puts it, engender some humility in those that think we understand how the brain works. If J&K are right, our techniques are not even able to suss out the structure of a relatively simple circuit, which, in most ways that count, should be more easily investigated using these current techniques. We can after all lesion a circuit to our hearts delight (but this does not bring us “much closer to an understanding of how the processor works” (5)) and take every imaginable measurement of both individual transistors and of the whole processor (but this does not give “conclusive insight into the computation” (6)) and do full connectivity diagrams and still we have little idea about how the circuit is structured to do what it does. So, it’s not only the nematode that remains opaque. Even a lowly circuit won’t give up its “secrets” no matter how much data we gather using these techniques.
This is the negative result, and it is interesting. But there is a positive observation that I want to draw your attention to as well. J&K observes that many of their measures on the processor “are surprisingly similar” to those made on brains. The cog-neuro techniques applied to the transistors yield patterns that look remarkably like spike trains (5), look “quite a bit like real brain signals” (6) and “produce results that are surprisingly similar to the results found about real brains” (9). This is very interesting. Why?
Well, there is a standard story in Brainville that promotes the view that brains are entirely different from digital computers. The J&K observation is that a clearly digital system will yield “surprisingly similar” patterns of data if the same techniques are applied to it as are applied to brains. This suggests that standard neuro evidence is consistent with the conclusion that the brain is a standard computing device. Or, more accurately, were it one, the kind of data we in fact find is the kind of data that we should find. Thus, the simple minded view that brains don’t compute the way that computers do, is, at best motivated by very weak reasoning (IMO, the only real “argument” is that they don’t look like computers).
Why mention this? Because, as you know, there are extremely good reasons provided by Gallistel among others that the brain must have a standard classical Turing architecture, though we have no current idea how brains realize this. What J&K shows is that systems that clearly are classical computational systems in this sense generate the same patterns of data as brains do, which suggests, at the very least, that the conclusion that brains are not classical computers requires much more argument than standardly provided.
At any rate, take a look at J&K. It is a pretty quick read. Both its negative and positive conclusions are interesting. It also outlines a procedure that is incredibly useful: it always pays to test ones methods on problems that we know the answer to. If these methods don't deliver where we know the answer, then we should be wary of over-interpreting results when applied to problems we know almost nothing about.
I think there's potentially another interesting analogy here between neurons and transistors. A typical computer (say a smartphone) has a very large number of transistors functioning essentially as switches and a much smaller number functioning as amplifiers. Each individual transistor is a very complex thing in terms of its physical properties. However, it’s only for the tiny fraction of transistors functioning as amplifiers that any detailed understanding of their physical states is necessary for an understanding of the whole system. I wildly speculate that the same may hold true of brains. What you really need is probably a high-level simulation of 99% of what’s going on and a much more detailed simulation of the remaining 1%. But you can’t know in advance what that 1% is. In other words, you can’t simulate everything at the maximum required level of detail, and yet you can’t know where it’s ok to skimp on the details unless you already have a pretty good idea of how the brain works.ReplyDelete