Via David Warren
at the Ottawa Sun. (I'm somewhat ashamed that I didn't catch this more directly!)
A new form of electrical component has been developed in semi-practical form. It's a 1971 idea, or more precisely God's idea finally re-manufactured. HP made the development; they're calling it a Memristor.
In simplest terms, it's a neuron.
What are the needed properties for a neuron? Its response - the amount of output it gives in response to each input - depends continually on the changes
it has experienced. I'm emphasizing changes
because everything in a nervous system responds to change, not to present state. Adaptation is the key. Your skin gets used to present temperature and notices a chill draft; your eyes get used to one shade of color and notice when it changes; your ears get used to the exhaust fan and wake you when it stops. Your visual memory gets used to the position of your furniture and signals you sharply when something is missing. Workers get used to a certain ratio of income to expenditure, and become bitter when it decreases. It's all
about rate of change.
The Memristor has exactly the right behavior. Its resistance - the amount of output it gives for each input - depends continually on the changes, the 'history', of currents that have flowed through it.
We've been able to rig up complicated circuits or software to imitate neurons, or nets of neurons, for quite a while, but the new Memristor has the needed set of properties in a continuous
way, at any point on a film or surface. This should make it possible to build a soft network with an essentially infinite number of paths. The film itself is surprisingly simple, nothing more than an extremely thin layer of titanium dioxide, the stuff that makes powdered donuts white.
A binary computer has to work very hard to imitate continuous variation: it has to run through a sequence of numbers, or use arithmetic to locate the correct number.
An innately continuous (analog) device can run through a smooth increase easily, but it has no convenient way to remember the direction
of the sequence. Again there are fancy ways to get there, but they tend to reach the same complexity as a binary computer, say 1000 internal components to hold one simple up or down curve.
One 'unit' of Memristance won't serve as a true memory; it will need to be combined and connected in various grid-like ways, and it appears that the HP researchers
are just beginning to travel that road. In order to reproduce brainlike functions, the 'wires' will need to have a huge set of potential connections, and the outputs and inputs will need a way to gather or perceive which areas of the grid are most active at each moment.
THIS IS BIG. Maybe not quite as big as Fire or Wheels, but pretty close.