Monday, December 02, 2013

Random Thoughts on the Virtual Human

kw: musings, artificial intelligence, computers, supercomputers

In one of Isaac Asimov's early Robot stories, the president of U.S. Robotics keeps demanding a more and more perfect robot, until at great cost a robot is produced that cannot be distinguished from a human. At that point aliens arrive, and in the course of time visit U.S. Robotics, where they are shown the robot. One of them turns and says, "So what is the point?"

What, indeed? It was published earlier this year that the largest supercomputer complex currently in use has a storage capacity and parallel processing speed that exceeds that of the human brain. What is required to do this? The facility fills a room the size of a medium-size warehouse and uses a total of 7 million watts, both for processing and for cooling. Of course, we are told that Moore's Law ensures that all this will "soon" be scaled down to a much more manageable size and power requirement.

I wonder what "soon" means. A loose way to state Moore's Law is that circuit density doubles about every two years. The Watson supercomputer, which has considerably less capacity than the brain—but is really, really good at looking up trivia needed to score points in Jeopardy—fills a dozen racks and would barely fit in a large bedroom. The warehouse-sized machine mentioned above must occupy 20 to 50 times that volume. Let's say, 2,000 cubic m, or 2 billion cc. The brain's volume is less than 1,500 cc. The ratio is about 1.33 million. Dividing in half enough times, we find it'll take just over 40 years for the brain-sized supercomputer to be possible.

Is that really needed? Why do we want to replicate humans? It presently takes 15-20 years and anything from a few thousand to a quarter million $US to produce an adult human, depending on education level and the level of development of the country in which this person is born and raised. No, the real desire for producing "artificial intelligence", aside from the coolness factor (yawn), is to have a machine that can do jobs humans don't want to do, or cannot do, yet requires more intelligence than we find in waldo-sorta robots (actually teleoperated devices).

At present, even Watson and the larger supercomputers are horribly inefficient. What would it take to do brain things the brain way, using silicon chips and wires instead of neurons and their support systems? Signals along sensory and motor neurons in the body are rather quick, in the range from 20-100 m/s. But they have a larger diameter than brain neurons, so the latter are slower: 3-5 m/s. Thus, it takes a signal about 40-50 ms to go from one end of the brain to another, such as from the optic lobes to the prefrontal lobes. This may underlie the frequency of the Alpha wave, at around 10Hz (8-12). It is the highest frequency rhythm in which the brain can participate as a whole.

But if you use coaxial cables or optical fibers or other kinds of wire for signal transmission, what is needed to keep signal transmission below 50 ms? Signal speeds in wire or fiber are about 2/3 c, or 200,000,000 m/s. In 50 ms, a signal can traverse about 10 million m or 10,000 km. So a mechanical brain can be continent-sized, leaving plenty of room for the really massively parallel kind of computation that the real brain performs.

A neuron is not just some on/off switch. It does some nonlinear processing of its own, so it would take a small CPU to emulate its activity (We'll need to learn a lot more about the various ways neurons in the cortex, the hippocampus, the amygdala and so forth, react to incoming signals).

Well, the brain has 10 billion neurons and 100 billion supporting cells called glia (of several different kinds). We can consider the glia as infrastructure and focus on the neurons. The modern package of a CPU is the cell phone. A smart phone's volume is 80 cc. 10 billion of these comes to 800 billion cc, or about 40 of those warehouses mentioned above. If they are "rackified" outside their cases, and powered externally so they need no batteries, and the radio module replaced with a hard-wire fanout, the volume can be reduced by 10, but you do need to wire them together. Each neuron has from 1,000 to more than 10,000 connections to other neurons, so it'll require a lot of wire. Still, we are in the range of a warehouse with less than 10,000 cu. m, quite a lot smaller than the continent we could fill, and retain the speed we need.

A question might arise at this point: Why is this system so much larger than one which is already faster than a brain? It is because I chose a neuron analog with around 1/10th the volume of a cell phone. The big supercomputer does not use virtual neurons; it does processing by a method entirely different from neuronal activities. But if you want to emulate the brain's functions, you have to emulate the way it does them. That vaunted supercomputer is not nearly as effective as the average house cat at recognizing faces, or voices, or footstep rhythms.

OK, whether it needs a few thousand cubic meters, or ten times that much, it might be costly, but it is theoretically possible to build a system that does brain things in the brain's way. But such a mind would be insane from the outset. Our brain is part of a complex system that includes a body full of sensors and an endocrine system (the original brain) with at least a couple dozen signalling molecules, and a few ways of expressing itself to external beings. Unless a brain, natural or built, has lots and lots of input and stimulation, it "spins its wheels" into helpless catatonia. Of course, using wires or optical fibers allows fast reflexes across pretty large spaces, but you'd really want a body much more human sized, attached to a brain locked away somewhere, via a fast data link no more than a few dozen km in length. A world traveler this one could not be. Remember, motor and sensory neurons have transit speeds in the 30-100 m/s range, so a 2m axon is traversed in 10-40 ms, but the shorter ones (upper body) on the fastest reflex arcs can get a signal to or from the nearest ganglion in a couple of milliseconds.

That's enough for now. And I still await my own criterion for genuine artificial intelligence: A mechanism, unaided, does its own research and development and obtains a patent.

No comments: