Four Paths to AI
Posted by Jeriaska on May 19th, 2008
I’m Jon Connell. I’m from IBM–yes, we still think about AI. My co-author is Ken Livingston. He is from the Cognitive Science program at Vassar, which is nearby in Poughkeepsie. What we are going to talk about today is four paths to AI–a little bit different from the other talks. I am not here to give you a prescription of what to do or what not to do. I am going to give you a description of what is out there.
What are the four approaches? How do you achieve AI? We are boiling down a broad series of approaches, and we feel there are four things people bet on–why are you going to win with whatever you are doing.
One of them is a silver bullet approach. ‘Everything is pretty good, we are just missing some special ingredient.’ Other people say, ‘You guys are so stuck on your mechanisms and your algorithms.’ What you really need to do is to figure out what is the important thing that it has to do. Does it have to emote; does it have to have conversations? What is that core value? If you get that core value right, then you can fill in the rest of the details–everything else will work itself out.
Then there is emergence. We got the conceptualization right, and all the mechanisms we need. We just need to pump in lots of something: compute power, data, experience in the world… something. Then, magically, we will have AI.
Then there is emulation. You have mechanisms and values, but you can still be completely wrong. Maybe we should take things that actually work in the world, like humans and animals, and start copying them at some level of detail. If we can make a good enough copy, we do not even have to understand it–it will just work as the original thing did. We can understand it later.
Let me give you some examples of this. Silver bullets: most of the technology is there, but some is missing. Then they all disagree about what actually is missing. If you pursue this kind of methodology, looking for the secret ingredient to make the complete recipe, you have to watch out because maybe you pick the wrong hole to fill. Silver bullets are great for killing werewolves, but they don’t do anything against vampires.
What are some of the silver bullets? Some people say fancy logic. First order logic is pretty good, but we need to use second-order logic, or epistemic, or autoepistemic, or some variation of that. Once we get the right core engine going, everything else is going to be fine. You can use all your vision stuff, your robotics, language parsers–but we need this logic at the core. That’s their silver bullet.
Another silver bullet is embodiment. You are never going to get the thing to understand physical properties like mass, velocity and weight, without a body. So, you have to have some kind of body, whether it is real (like a robot), or virtual (as it is in games), but you have got to have that component to your system.
Another silver bullet: quantum physics. There is a mystery in quantum physics; there is a mystery in intelligence… you link them up, and maybe they are the same mystery. God lives in those little Heisenberg uncertainty gaps. You have to have quantum at the bottom, otherwise the ghost cannot come in and haunt the machine properly. You will never have a soul in your machine.
Another silver bullet is inexact reasoning. Logic is approximately right, but, man, it way too brittle. Anytime it hits the real world, it tends to break. You either haven’t modeled something right, or you cannot measure it accurately enough, so we really need to be able to handle these shades of gray. Even in medical expert systems you see this sort of thing.
Fuzzy logic–you can describe something in what seems like decent English and then have a smooth control surface result from that. The fuzzy nature of it is kind of nice; it accommodates the bad sensors. If you have worked with robots you know that you are never sure if the wall is really there or two inches over. You can also do some blending of control actions. That is one of the big things with robots–choosing which action to pick. Sometimes you have to trade off a lot of different things. There is a lot of fuzziness there, not just “do action A or do action B.” That falls nicely into inexact reasoning.
Another silver bullet is deep language. So much of being human is tied up with language. If you do not have language, no one is ever going to think that you have built something that is human. It is also a great way to learn, because there is lots of written material. Our kids learn by looking at written materials, and teachers feed them materials that are written. An old program that did this is Dyer’s BORIS program, which I still think is one of the best natural language programs. It is limited in its scope, but what it did was pretty cool.
It has lots of levels of understanding of the text, and various data structures for types of understanding–temporal understanding, emotional understanding, spacial understanding.
This is the kind of thing it can do. It reads this little paragraph about these two roommates, then it answers questions from it. “What happened to Richard at home?” One never said he was “at home,” but you know that letters typically arrive at people’s houses. You can figure that stuff out. You look at these answers and figure this computer really understood this paragraph as good as humans. In some ways, this is very convincing AI. Unfortunately, for this particular program there was a lot of handcrafting. If you could figure out a way for it to auto-self craft that would be nice.
Those are a sampling of some of the secret ingredients, or silver bullets, out there that people talk about. Another category of approaches is to not focus on mechanisms, but focus on what you want to end up doing. Then, what do you want it to end up doing? Is it very important to have an emotional computer that breaks down and cries sometimes when you tell it to do something hard?
The problem with core values is what if you pick the wrong core value. There was this theory about the Four Humours which held sway in the medical field for nearly two thousand years. The way a fever would be cured is letting out blood. You have excess hot humour, and you get rid of it, you figure it will go away. It kind of did not work out that way.
Some of the core values include situatedness. It matters that the robot have some consequences for its actions. It has to be in the world and able to interact with the world. You cannot just have a box that sits there and answers questions sporadically. It would have to have its own desires.
Another one is hierarchy and recursion. You can build neural networks that do simple situation-action rules, but how does it ever do something more than that? How does it build successive abstractions on top of the next abstractions? You really need this way of making abstractions at the core of the system, since that seems so very human–animals, for instance, do not do it very well.
Another core value is self-awareness. If you ever remember the bomb from Dark Star, he argues with it, “Your purpose in life is to explode.” Finally, the bomb decides it’s okay to explode. That’s the important thing–you cannot really be a human-level AI unless you understand that you are a human-level AI. We should focus on that, and if we get that right, then all the mechanisms we have in the corners will be suitable.
Here is a big one–emotion and motivation. Emotion is not something you add as a user interface to make the system more friendly. It actually is some core aspect of the system. As much as inexact reasoning handles the fuzziness of the world, this handles some of the value structure of the world. How do you decide what to remember? How do you choose which action to do? How do you decide if a place is a good place to be, or a person is a nice person to ask a question of?
This is very important, and it shows up in some places that you would not necessarily think of as emotion. Stewart Wilson had this thing about exploration policies and reinforcement learning. For a system to learn, it has to sometimes not just try to optimize its goal but goof around in the environment and figure out how it works. How do you figure out how much time to spend goofing around? You can use a small fraction of the time all the time, and then use what you have learned. So, I really am hungry and have to go get the food, but sometimes I will turn over rocks to see if there is anything underneath them. Another policy is to explore a lot at the beginning, then less and less over time. Children play a lot and adults really do not play much. When you have some expected optimal performance–if you are not hungry most of the time, then do not bother exploring anymore, because you are good enough for your environmental niche.
It goes on. He has about ten policies. Some of the state-based policies are interesting: If your predictions are failing, then explore more. If you are confused or not learning very much, then you do some more updating. If you are in a situation where you are learning something, it is probably a good environment for continuing to learn, and you should learn more there. Then there is a boredom aspect. If you have nothing better to do, then start poking around.
Emergence–everything is already present, it is just a matter of scaling the system up. What fodder needs to be provided to this engine, that is where the approaches differ. The problem with this is that you have to look out for is early asymptotes. A battery will make a frog’s leg twitch, so if we put a lightning bolt in there, maybe the frog’s leg will get up and dance around… It didn’t work.
Axiomatization is where you build a logic formalism, but you say you need to axiomatize time, space, and liquids… there are a lot of things you could write down axioms for. There is the approach of evolution–basically you just give it lots of cycles in some environment. That is what you are giving it lots of, along with some general rules for improving itself. Integration is another form of emergence. You have to keep throwing lots of different components together. It cannot work unless it has auditory perception, or without olfaction it will never go, or it needs dexterous manipulation. Once you have all those things together, maybe everything will work. That is another kind of emergence.
There are lots of facts–Minsky says a couple million. This is an example of some of the facts you might have in your head. Reinforcement learning is another one where you need lots and lots of data. This is not just passive data, but active data, where robots actually explore. This is a simple Pac-Man environment, but they allowed it to explore this environment for 15,000 steps. If you are actually doing that with a robot, 15,000 steps takes a lot of time. Basically, experience is what you are feeding it here.
The fourth category is emulation, which is, everything you know is wrong–just start copying stuff. But what to copy, and what level should we copy at? You might want to start copying details that did not make any difference at all–we don’t need feathers on our airplanes. Maybe they would be better if we had them?
Various emulation levels–we need to simulate right at the neuron level. Or, we need to do neural networks. We understand basically how the neurons function, but we need to understand the networks, so we will just copy them. There is a lot of work on this kind of thing. Some say human development. Children develop from knowing basically nothing to being full humans. How does that happen? Maybe we should copy that process.
Humans evolved from animals. Maybe we can learn something not from human-level AI, but animal AI. Then it will be a little step from there to human AI. For instance, you can teach something like rat learning. Sociality says that you have to actually talk to humans and learn from them. The more you talk to them, the more humanlike it will become–sort of a cultural transfer. Kismet is one example of that, designed to interact with humans.
What good is this classification scheme. We broke down the world into these four mystical thoughts about why AI is going to work. What good does it do? Sometimes there are commonalities within the approaches that are interesting to look at. For the silver bullets, everyone agrees that there is some kind of substrate underneath, so what is that substrate? Is it symbolic, set-based, or fuzzy? A lot of the core values seems to revolve around language and sociality, if you look at some commonalities there. Maybe that is useful.
The big thing with emergence is that it is a scaling phenomenon. Is there anyway to guess how much data you are going to need? Is it a practical amount of data. This asymptote, which is the Achilles heel of the approach, is there any way to tell if it will asymptote by watching your system performing. It gets better and better at first, but is it going to slow down? Is there any way to predict if that is going to happen? Some common questions for emulation–what is faithful enough? Do you have to get down to ion channels? Do you have to get down to chemistry or the quantum level? How far down do you have to go to be faithful enough? Again, are there any common principles that show up in this emulation? Auto-encoders seem to be fairly popular in the neural community. A lot of people say entropy reduction and reinforcement are important.
The ten thousand foot view of why this is important is that it helps to defuse some arguments. When you say, “Your system sucks,” are you objecting to the technology or its methodology? Are you saying, “Your inexact reasoning engine is never going to pan out.” Or are you saying that it is not one little piece that is wrong, you need to go in a whole different direction? I think of this as the pacifist talking to the hawk. The knee-jerk response is “A gun won’t solve any of your problems. Why do you want to be so violent?” But if you assume the other guy’s methodology for a minute, if you are going to fight, at least a gun is useful in that situation.
Then sometimes there are some cross-methodology issues. I talked about commonalities within the approaches. A lot of these things seem to need feedback and language at some level. We have these four broad categories, is any one of them going to really win? You might have to combine some of them, like a core value and a silver bullet. Self-motivation is a core value, and language is a silver bullet. Or maybe emulation and emergence together will do it. Neural networks emulate the brain, and then we will feed lots of web data in to provide emergence. There could be others.
I have my own personal view. There are these four categories, and maybe we can learn something by looking at the structure of those categories. Thank you.
