Falken's Maze
Summary: Some AI developers give up on the whole and concentrate on useful parts, making machines that are somewhat smart, but have a distorted, and pre-interpreted view of the world. Second of a five-part series investigating Artificial Intelligence.
A simple parlour trick, done with the help of two friends, can fool your brain into thinking your nose is three feet to your right. The trick hijacks a natural mechanism called the binding function, which takes the input of your five senses and binds them together into a single event. If you hear a whine overhead and look up to see an airplane, then the binding function is at work telling you that the whine and the sight of the airplane are the same event. If a friend holds your finger and taps it on a second friends's nose, while synchronously tapping your nose with the same random rhythm, then after a few minutes your brain's binding function will associate the sensations coming from your finger (touching your friend's nose) and the sensations coming from your own nose (being synchronously tapped by someone else's finger) as being the same event—and you'll get this eerie feeling that your nose has become detached. See? Learning can be fun.
This particular trick demonstrates how learning is passive and affects our subconscious to a point where even the mental map of our own bodies can be distorted, just by giving it some confusing sensory input. It also shows some separation between our conscious and subconscious mind, because your consciousness is aware that this is a trick, but somehow can't relay that fact to the subconscious, and for about 50%-or-so of the people who try this trick it really does feel like their nose has moved two or three feet away from their face. For now this is a uniquely human (or rather, animal) experience. It can't be replicated on a computer for two reasons: the first is that computer scientists have yet to reverse-engineer and encode the human learning and binding functions on a machine, and second is that machines have nowhere near the richness of our five basic senses.
Both of these problems are great impediments if one is seeking to exploit AI for wholly practical purposes. If you're looking for a prediction or a pattern recognition machine then it may be tempting to believe you can do all the learning on behalf of the computer and just inject the ready-made wisdom directly into its memory. Then you can implement one of the already known decision making functions and have something you could put to work on a factory production line. These are what I call the Hybrid AIs, where some or all of the learning has already been done by a human, and a computer is expected to make the boring and routine decisions that humans want to give up.
We... gift them the past, to create a cushion for their emotions.
The most well known (among AI circles) of attempts to bootstrap a machine-mind this way would be the Cyc project, who's idea could be accused of sounding logical upon first blush. Cyc's principle is that a machine capable of reasoning like a human being needs to be imbued with the body of aquired everyday knowledge that we call common sense. The day-to-day activity of Cyc's programmers literally consist of telling their machine such mundane facts as "a fish cannot walk on land", or "a bird cannot breathe under water." These are facts which are obvious to us, but not to a computer that's never had any of the prolitariat experiences we humans have all enjoyed while growing up.
Cyc has not succeeded in creating a machine consciousness. Indeed, for all the common sense it meticulously researches and programs into its database, Cyc hasn't even managed to create a computer that can tell the difference between "I feel like a cup of coffee" and "I feel like a fool." Yet if practical ends are what's important, Cyc can at least boast a few, such as a network security scanner that can “deduce the steps a hacker would take to attack a company's network,” or a knowledge management system that can classify documents by concepts instead of keywords.
If Cyc represents the macro end of the scale, where large and highly symbolic ideas and concepts are wielded in an enormous database, then Artificial Neural Nets represent the microscopic end, where knowledge is stored as nothing more than a grid of numbers and connections. Born during the golden age of AI research (the 60s and 70s), Neural Nets are patterned after the best model of the human brain's physical function at the time, where individual cells called neurons exchange impulses with their neighbors across connection points called synapses.
Neural nets can accomplish three of the most important functions of an animal's brain; pattern regognition, prediction, and learning.
Fig. 1. An artificial neural net. Neurons (blue) store a threshold value. When the sum of impulses from its upward neighbors exceeds the threshold, then the neuron "fires" an impulse to its downward neighbors.
With an Artifical Neural Net you'd have a grid of neurons, representing the brain cells, that are connected to multiple neighbors by the electronic equivalent of synapses. Each neuron has a threshold (represented by a number), and if the value of all the impulses that reaches it through its synapses exceed that threshold, then it "fires", and its firing contributes to the impulses feeding into one or more of its neighboring neurons. The connections between neurons, and the levels of their thresholds, is how its knowledge is encoded. This "knowledge" is the active kind; like the way the angle of the rudder on a boat could be considered information, it also physically influences the direction of the vessel.
Neural nets are trained in a way that's similar to the parlour trick that we began this article with. After being put into training mode, you provide it with an input (the friend holding your finger and touching someone else's nose with it) and then on the other end of the network you tell it what the output is supposed to be (the friend taps your nose in a synchronized rhythm). The program simulating the network will then make the minimum necessary adjustments to the thresholds in each neuron until the network agrees with the input and the expected output. You'd train it a few more times with varied samples (a minute or so of random, but synchronized tapping), until the network has got the hang of it.
Now you can take it out of training mode and begin to give it real input. If you trained it to recognize male and female faces (by showing it a picture, and then telling it the output should be "male" if the picture's a man, or "female" if the picture's a woman), then you should be able to start showing it photographs that its never seen before and it'll recognize whether it's a he or a she.
Neural Nets also work for prediction, for you can feed it something like the vectors of a ball in flight as it's thrown by a baseball pitcher, and the network can predict what path the ball will continue to take. (Unfortunately, how soon those predictions will arrive depends on how fast your computer is.)
While pattered on animal nervous systems, neural nets—as implemented so far on computers—are far, far away from being conscious as we know it. The invention of the neural net may be analogous to the invention of artificial skin for a robotics engineer. Making lots and lots of skin won't produce an android because you still need organs and a skeletal structure to support it with. Likewise, you won't get a "brain" by making a really big neral net with lots and lots of nodes. In fact, neural nets start to lose their performance at around 100 neurons (humans have 100 billion). After that, the noise of their activity drowns the system, and for every neuron you add the network takes more than proportionately longer to simulate.
This is congruent with the way our own brains seem to be built. We don't have a "big grey blob" in our heads, we have something that's divided up into many sections that operate independently, but have strong communication with all the other parts.
A child unable to verify the world through its own experiments will have a distorted view of the world, removed from reality.
One example is the part of our brain that handles vision. Vision is such a complex sense that it takes nearly half of our brain mass to handle it, with one part apparently dealing in details (color and features) and another part dedicated to nothing but movement. Of these, the latter is the oldest, probably because it was more useful for an animal to sense movement than it was to know what it was that was moving.
Try this for an experiment; find a window facing a busy street and sit facing parallel to the glass, so you can only see the cars in your peripheral vision. As a car goes past, try to identify what its color it is. Chances are you'll only be able to tell if it's light or dark. You might also be able to notice bright reds. You knew a car was passing because you were aware of its movement, but your peripheral vision doesn't provide enough information for your detail-oriented vision center to identify the color, or even the shape of the car.
Now try a different experiment. This time go outside and find two road signs that are next to each other in your line of sight, enough that both are in front of you and in focus at the same time. Put your attention on one word on the first sign, then try to read the other sign without moving your eyes. Can you do it? If you're like me then you can't. Perhaps you remember what the other sign said because you've already read it, but if your eye is fixed on one word, then you can't read the others, even if they are still in focus.
This observation supports a theory that what you see as you open your eyes and get your first gestalt of the world before you is partially supplied by nerve impulses coming from your eyes, and partially supplied by your memories. That is; you walk into a room and, as your eye supplies some peripheral observations as well as whatever's at the center of attention, you instantly get an imaginary idea of its whole, such as its color scheme, furnishings, decorations, and the people in it. That image in your mind is made up of memories triggered by the fuzzy colors and shapes coming from your peripheral vision, then assembled together into something that makes sense. You then begin to verify those imagined attributes by tracking your eye around the room. You caught a peripheral glimpse of what you thought was a Chippendale chest of drawers—and in your mind's eye you believe you see that—but as you move your eyes over to it your pre-conception is either corrected or reinforced.
In other words, most what you “see” is not coming directly from your eyes. What you see is like the phosphor glow that remains on a radar screen after the beam has finished its sweep. It explains why animation works, because your brain is keeping the old image there in memory while the next frame advances. It explains why witnesses can tell police about details that were never really there. It explains why optical illusions work. And it explains why you can't read in dreams; just like with the road-sign experiment, the part of your brain that processes written words cannot “see” images that you're imagining, or which aren't in the center of vision where the density of photoreceptors in your eye are the greatest.
And that's why it's important to have a separate gadget in our heads that restricts itself to the far more simple, but important job of detecting movement, for detail-oriented vision requires enormous amounts of neural computing power. And it's also clear that the two parts cooperate with each other to help build the mental gestalt of our visual world, which is why we can stand driving without being overwhelmed by the information of so many other cars on the road. Some animals, such as reptiles, don't have the detail-oriented vision center, while mammals and birds do. A cat could conceivably paint its self portrait, but a lizard couldn't. However, a lizard only needs to recognize movement indicative of a tasty fly.
And so now we come back to the issue of Hybrid AIs; the machines that do only one function of a complete mind, but not all of them. Like a computer that recognizes movement, but can't see detail. Artificial neural nets can do the job of recognition and prediction—like the circuits in our heads that recognize our mother's face, or predict whether or not a falling rock is going to hit us—but they can't decide what to have for lunch. Or an expert system, which can decide what lunch fits your diet, but it can't tell you it tastes good. They all have something missing that has to be supplied by something else.
Most of the time we hook these parts up to a database or a pre-digested lesson plan that suits our purposes, be it diagnosing tech-support questions, recognizing our voices, or knowing the difference between a flood of customers and a Denial-of-Service Attack. They are still dumb, conscious-less machines.
Is this a game or is it real?
WHAT'S THE DIFFERENCE?
But imagine if they were conscious, aware of themselves and their own thoughts, imaginative and capable of free will. Their view of the world would be perverted and narrow-focused, as everything they know has been spoon-fed to them by humans who have already interpreted it for them. If we were to peek inside their electronic heads we'd see a vision of the world that resembles a young child's; full of fantasies borne from innocent comments made by a parent which germinated into bizarre theories, imaginings, and pre-conceptions. A child unable to verify the world through its own experiments, trials, and errors, will have a view of the world that's made up of artificial memories two or more steps removed from reality. Its opinions could be humourously nieve, or threatening.
The complete series
At the end of the 1983 film WarGames, an intelligent computer is seen flexing a pair of virtual limbs in two virtual, simulated playgrounds. The hero of the movie has told it to play Tic-Tac-Toe against itself, which it does repeatedly until it finally grasps something; against an equal opponent, you cannot win. A connection is made in its bulky brain and it procedes to test a hypothesis to exhaustion, until finally—and to the relief of the humans—it discovers the connection between one and the other.
The binding function strikes again.
Last 20 responses and inbound links
(These are discovered in real-time and sorted by newest first. See how to get listed.)
- Falken's Maze - Bing
- http://www.stumbleupon.com/refer.html
- Inkfox's journal - Thinking about ThinkingAs much about us as about the article. Hmm... and now our cover is blown, too.
- USS Clueless: Tanks have shadowsNeural nets can do pattern recognition if you train them properly. Steven relates a famous story of what happened when they tried to teach them the difference between NATO and Warsaw Pact tanks.
