Slashdot Mirror
Researchers Build Evolving Brain Computer?
destinyland writes "'We have mimicked how neurons behave in the brain,' announces an international research team from Japan and Michigan Tech. They've built an 'evolutionary circuit' in a molecular computer that evolves to solve complex problems, and the molecular computer also exhibits brain-like massive parallel processing. 'The neat part is, approximately 300 molecules talk with each other at a time during information processing,' says physicist Ranjit Pati of Michigan Tech. When viewed with a scanning tunneling microscope, the evolving patterns bear an uncanny resemblance to the human brain as seen by a Functional MRI. Using the electrically charged tip of a tunneling microscope, they've individually set molecules to a desired state, essentially writing data to the system. And while conventional computers are typically built using two-state (0, 1) transistors, the molecular layer is built using a hexagonal molecule, and can switch among four conducting states — 0, 1, 2 and 3, suggesting it may ultimately have more AI potential than quantum computing."
Goddamnit, that is not how it works. Even if each molecule has four different states, you can easily map them onto a small, finite number of bits - you just represent each molecule with two bits in a computer, and there's your equivalency. You don't get anything out of more states per unit except higher density. Seriously, TFA doesn't make this mistake; why did you have to add some useless speculation to a perfectly reasonable article?
Modeling neurons inside the computer is how people have been doing it until now. And while it has made steady progress, it hasn't proven terribly successful; since the advent of the computer age, these AIs have evolved from being equivalent to a flatworm to being equivalent to a guppy (and I'm being optimistic here). Trying to model a massively parallel process inside a serial computer is not terribly advantageous - scientific computations such as CFD (computational fluid dynamics) and MD (molecular dynamics) are plagued by the same limit. What we really need for these kind of processes is a computer made out of very simple, small and fast elements that do exactly the task you want them to do and that are all connected. There have been steps in this direction (earth simulator, GPU computing,...) but I feel the current approach can easily trump them all - at least for the purpose of creating AI. Scientific calculations will be another ball game, because there, the desired properties of the system are very rigidly defined.
This is not to say there is no room for classical computers - some problems are inherently discreet and serial, and there, our serial processors rule. At least until quantum computing becomes more mature ;)
And while it has made steady progress, it hasn't proven terribly successful; since the advent of the computer age, these AIs have evolved from being equivalent to a flatworm to being equivalent to a guppy (and I'm being optimistic here).
And how many millenia did it take for the biological process to create a guppy from a flatworm ?
Considering everything we've achieved in the last 50 years, I think the next 50 will be even more revolutionary for the AI overlords using us as 9 volt batteries.