Slashdot Mirror
Neuroscience Can't Explain How a Microprocessor Works (economist.com)
mspohr writes: The Economist has an interesting story about two neuroscientists/engineers -- Eric Jonas of the University of California, Berkeley, and Konrad Kording of Northwestern University, in Chicago -- who decided to test the methods of neuroscience using a 6502 processor. Their results are published in the PLOS Computational Biology journal. Neuroscientists explore how the brain works by looking at damaged brains and monitoring inputs and outputs to try to infer intermediate processing. They did the same with the 6502 processor which was used in early Atari, Apple and Commodore computers. What they discovered was that these methods were sorely lacking in that they often pointed in the wrong direction and missed important processing steps.
Anyone with even an elementary education in cognitive science will tell you that attempting to model thought processes is always done according to the dominant technology of the time in question. First it was machinery, then it was circuits, then it was computers.
This does not mean the model is accurate or even useful.
In order to understand the DNA of an Orange, we "scientists" dissected an alarm clock. This _proved_ that our methods of studing oranges, and fruit in general, have been wrong for centuries.
I know I'll catch hell for my religious beliefs, but...
I think that the 6502 was not the result of evolution, but rather it had a Creator and was the product of Intelligent Design. There are just so many subtle clues that suggest features that were deliberately put in there. Could natural selection really explain how it had two different indirect access modes, one that selects a direct index from an offset, and the other adds the offset to the index?
These researchers may be trying to apply the wrong methods to a device that is almost certainly the product of a higher power.
A dingo ate my sig...
We once had electronic circuits designed to perform calculations (example: Enigma). It seemed natural to try to model the brain as a complex electronic device.
We now routinely use silicon integrated circuits to perform calculations (example: the IBM PC-XT). It seems natural now to try to model the brain as a complex general computing device.
The take-away point I get from this is that we may need another revolutionary technology or two (fully three-dimensional integrated circuits? IC's based on carbon instead of silicon?) before we can model the sentient mind as similar to an artificially created device. Such advances may also be required before we can create (invent?) a true "artificial intelligence".
Neurons aren't digital processors. A set of connected neurons isn't either. Neuroscience already knows that it's really difficult to learn about the structure and function of the brain from the available tools. What was more interesting was that they were able to pick up anything. They found that the chip had a master clock, for example.
There are people already challenging the use of viewing the brains as a computer (signal ins and outs) in terms of really understanding how brains organize and function. So, given all this, it's not surprising that the methods didn't fare well. The neuroscientists already knew they had a very tough task, it's those in CS and AI that are assuming that understanding the brain is the same as understanding a collection of digital circuits.
It's not brain surgery.
...a 6502 is not a brain.
The issue is, the 6502 is several orders of magnitude less complex than a brain. It could be likened to a massively parallel computer that is running thousands of programs all at once. So it is completely reasonable, on the scale of the brain, to suggest that damage to an area in a dozen people that affects their hearing to draw the conclusion that that part of the brain is responsible for hearing. Damaging a couple transistors in a 6502, a single processor, is akin to damaging a few neurons in each of a million different areas of the brain at once. The researchers correctly determined that a particular damage which caused Donkey Kong not to boot did not mean the part of the CPU which is responsible for Donkey Kong was damaged. That does not, however, say anything about the methods used to study the brain. It is, unfortunately, yet another reason why researchers need to stop making silly comparisons between computers and brains, and also need to stop "playing for the crowd", which really was what this seems to have been.
We don't know everything about how the brain works but we know a lot. How much is left to be known remains to be seen.
And that makes it a strawman, how? The general thrust of his argument is that the detractors of AI are so hung up on the concept of intelligence that they never address whether we will be able to artificially replicate the functions of the brain or how close we are, or what is in the works to get us closer, or anything in that area really.
Once we eliminate all the posturing around the concept of intelligence by changing the term to machine learning, all the arguments collapse.
What you're proposing is basically a GA: Genetic algorithm.
Even when you give a system a biological analogy as its base, the results are unpredictable, un-interpretable, and don't confirm to any logical architecture.
There is a famous example of a chip designed to detect two different fixed frequencies of an input signal, and output which is active (if any). Designing the chip by hand results in a working, logical model of a certain size.
If you allow GA to run random "evolution" over the circuit contents, punishing it when it gets it wrong, and breeding from it when it gets it right, you end up with a circuit that appears to do the job.
Ironically, it even does it inside a smaller space than the human would have designed it. However, trying to interpret HOW it does that job is almost impossible and certainly not worth the effort. But the problem is, if you want to USE that chip, you have to do that effort. One day, there might be a corner case where it doesn't operate as you believe it might, and you won't know until you hit it.
At least with a logic circuit you can understand, you can in theory mathematically prove what it will do quite easily. With one that has multiple feedback loops and randomly-built interactions between parts, analysing it isn't worth the money you'd spend doing so, especially as it's quite likely that even after millions of generations of training, it could still contain quite prevalant bugs (i.e. when exposed to a real-world frequency close to the target ones that fluctuates differently to how whatever training inputs were used).
And GA's have proven themselves not quite as useful as we first hoping. Millions of generations later, you can still fall flat on your face and there's no real way to steer things differently without doing it all over again, and no reliable way to understand or adjust the output in even the smallest way.
Whenever you see that an AI has been "trained", you should be suspicious. It's like saying a dog has been trained. It's still an unpredictable, ever-changing, free-thinking animal that we don't understand but which usually gives us the output we want (sit, stay, heel). There's no telling, though, when it might decide to turn around and bite you, because it's range of inputs is not the only factor in how it makes a decision.
And that's a model of a system that, generally, abides by rules, accepts training, etc. and operates in certain logical ways to ensure survival after millions of generations of evolution. Anything we fabricate has even less guarantees.
The 6502 is not a microcoded processor.
Contribute to civilization: ari.aynrand.org/donate