Friday, September 21, 2012

Artificial Intelligence and Cognition: Defending my Optimism

Visit the High Level Logic (HLL) Website
— Home of XPL (eXtensible Process Language)

It's like XML, but you can store source code in it.

“Engineers are naturally grumpy,” I once said. “If we ever start thinking those damned machines are something other than damned machines, we'll get packed away in white jackets.” Does our own natural attitude playing with our interpretation of progress have something to do with why artificial intelligence seems perpetually elusive?

I'm an optimist on developing artificial intelligence, but not very optimistic about convincing others that anything we do can be counted as much more than a trivial re visitation of things that have already been done. I think that engineering “grumpiness” is a key to understanding that no matter how far we get, accomplishing anything really interesting will still seem to be a long way off.

First let me say that I'm grumpy too. I can't help it. It's really more difficult for me to get a cuddly feeling from a robotic baby seal than it is for an old woman in Japan. I know too much about what's under the hood, or its skin? And my problem, which I'm certain is shared by other engineers and scientists, isn't just an emotional one.

My optimism also comes from knowing what's under the hood. My optimism regarding artificial (meaningful) self-awareness for example comes from the design of robots that learn and adapt. In a design by Peter Nordin (related article), robot software learns about the physical robot it runs by exercising its actuators and discovering their effects. The robot uses this self-knowledge to efficiently begin learning more complex behavior and ultimately about its environment and how to effectively interact with it. Instilled with “motives” their behavior becomes useful. It's also been demonstrated that robots can learn from direct human verbal interaction as a replacement for programming. (For example: English translation of Swedish documentary) Self-aware robots can distinguish between themselves and others and learn through interaction how they should treat others, a pathway to robot ethics.

“Humbug!” pronounced a world-famous engineering professor. (I paraphrase, perhaps very badly, to emphasize engineering grumpiness. I'll leave his real name out to allow my fictional character to more clearly illustrate the problem. It's “reality-based” we can say.) “In that last step,” a small demonstration of an additional idea at the end of a larger project to do other things, “rules were used. RULES! You can't get anywhere with rules. What they're trying to do is still decades away.”

It's not just world-famous engineering professors. As I've said, I suffer from the same affliction, as do many. I read (again) recently about a robot taking a self-awareness test that has often been given to animals. Place a robot in front of a mirror and see if it recognizes itself. “BAAAH HUMBUG!” I thought, seemingly without any ability whatsoever to restrain myself. That's just pattern recognition, no different than recognizing anything else. (The actual test involves changing something about the creatures appearance, typically by placing a mark on its body that it only sees in reflection, and seeing if the creature notices the change as being to itself, usually by pointing to it or touching it on its own body rather than in the reflection.)

Oh, but wait! Will this be the first time such an experiment has been conducted? If so, it would be a dandy wonderful experiment. It's the same one used by psychologists on living creatures. If successful, it in some way would prove that something interesting has happened (no matter how long we've understood that it could.). And in fact, I can actually use a phrase from the paragraph above, in which I was explaining my optimism, to explain the potential importance of this advance. Robots that recognize themselves can use that ability to “distinguish between themselves and others.”

But can we accept any of this as having anything really to do with self-awareness? Or are we (engineers at least) forever going to be the nay-sayers on account of the fact that we know the magic trick behind the “illusion”? I argue that we can be more positive and optimistic, and accepting of progress toward artificial intelligence and cognition by truly embracing our grumpiness. THEY'RE MACHINES!

Are we too often, unknowingly perhaps, making the mistake of thinking that developing artificial intelligence is synonymous with creating a synthetic human? Must machines hold the same mysteries of their inner-workings from us as those of living things in order for us to allow that we're well on the way to artificial intelligence? Must the goal always be hidden away in what we don't already know? (I could segue into the “singularity” here, but I won't … just mention it's where we expect to no longer understand what's going on.)

The trick, for us, I think, is to accept that machines are machines. They aren't something else. Let's extend the description of that latter experiment like this. A machine is let loose in a room with a mirror. It autonomously roams around the room and sees the mirror. Upon further investigation (still, autonomously), it sees its reflection and says, “Oh look, that's me.” That really is a pretty good trick for a machine. Does it wave its hand to confirm that it's looking at itself (perhaps only then learning from the image, how it looks)?

I don't have the sense that I've nailed this argument to the wall simply by posing the question in some context. As a proxy for things we already understand, let us reconsider that most humble of artificial intelligence techniques; rules. During commercialization of rule-based expert systems the 1980s, they were imagined as a step toward all kinds of software magic. These systems were after all, the product of artificial intelligence research. By the end of the 1980s, great expectations had crashed on the limits of those early rule-systems and it was thought that they should never be mentioned in this context again.

Now presume for a moment that I am a technically well-educated human being with experience. I understand how to use a rather wide range of techniques to solve problems, answer questions, and even to trigger decisions. Sometimes the use of one sophisticated technique, some kind of statistically analysis for example, is enough to accomplish what I need. Sometimes it takes a string of sophisticated operations to get where I need to go.

Many of the techniques I use can be performed with the help of a computer. In the example, statistics, there is much software available to support the task. Computers can already do this work much more rapidly and reliably than people. The trick here is to identify the right technique for each task, set things up, and run. I'm so smart. I'm human. I can do that and computers can't. Why? Because I have knowledge stored in my brain about what those techniques do. How do I apply that knowledge? Well … aahh … uhm … it's sort of like rules. Should I go ahead and develop a more autonomous level of useful computer processing, or just say naaaah, that would be so-o-o 1980s?

What if my more autonomous system could in fact perform this feat, but competes oddly against humans regarding which kinds of cases it understands how to handle? In other words, what if it has limitations, but they aren't the same as my limitations? In comparison, there are still some cases that I can handle better than the computer? Oh, yes, by the way: I have limitations. (Admit it, so do you.)