Justin Li

Musings on Human-Level Artificial Intelligence


I've been thinking a lot about human-level artificial intelligence (for convenience and humor, let's just call it HAI) lately. I suppose it started in November, when I was looking at Carnegie Mellon for grad school and I stumbled across Professor Scott Fahlman's blog, "Knowledge Nuggets". Although his research is in knowledge representation, he writes about HAI as well. In several posts, he outlined what we are missing in current AI research, and what he thinks a knowledge base (KB) for HAI would be like. Although I've never communicated with him, he was the main reason I chose to take Knowledge Representation this quarter instead of Introduction to Computational Linguistics. It reminded me that my real interest is in the artificial creation of a psychology. For a while I was distracted by other, perhaps much easier and more practical fields of AI like textual analysis, but reading Fahlman's blog brought back my interest in strong AI (an AI which can actually think, as opposed weak AI, which only gives the appearance of thinking).

Since then I have thought more about this problem, and I would like to use this post to organize my thoughts. Reading first Alan Turing's essays from the beginning of the digital computer and his visions of what computers can do, then Douglas Hofstader's Godel, Escher, Bach on symbols manipulating symbols, resulted in a large number of ideas. Comparing those ideas with the current state of AI, I also see some difficult problems to solve. I would love to tackle some of them, and while I don't think having HAI is impossible within my lifetime, it will definitely take a lot of smart people and clever innovations.

Let me start, then, with a quote from Turing, from his paper "Computer Machinery and Intelligence":

Instead of trying to produce a programme to simulate the adult mind, why not rather try to produce one which simulates the child's?

It occurred to me that a lot of AI research is done in replicating what we see in adults. This not only applies to early AI research, when theorem provers and chess masters were written. Planning, problem solving, knowledge representation, reading and understanding language... these are all behaviors which humans learn relatively late in life. While I don't doubt there are many practical applications of results from these areas - that may perhaps even be the reason why there is so much research - it seems difficult if not impossible to arrive at a general HAI from this direction. Intelligence itself is a complex enough creature; studying it after it has matured and grown is like trying to reconstruct a tree. Although recreating the tree may be the ultimate goal, studying the structure of the seed is the better path for research. A successful replication of the seed necessarily leads to the replication of the tree, and yet the seed is infinitely simpler than the tree with its myriad of branches and leaves and flowers.

Similarly, understanding the cognition of a child - or perhaps even an infant - might be a more worthwhile direction of research. Continuing from Turing's previous quotation:

If this [the child brain] were then subjected to an appropriate course of education one would obtain the adult brain. Presumably the child brain is something like a notebook as one buys it from the stationer's. Rather little mechanism, and lots of blank sheets. (Mechanism and writing are from our point of view almost synonymous.) Our hope is that there is so little mechanism in the child brain that something like it can be easily programmed.

Of course, I understand that while there may be little in the brain, that doesn't mean that it's not complex. It's simply because it's less than an adult brain that it's our object of study. A similar argument can be made for studying the brains and intelligence of other animals, and there would probably be contributions to be made there, but the gap between human intelligence and animal intelligence is too wide to only study animals.

Here, I would like to point out that studying the intelligence of humans or animals is not the equivalent of studying the brains of humans or animals. Studying the neurological processes in the brain to arrive at intelligence would be like building a car from quarks and electrons. Hofstadter writes:

... we hope that thought processes can be thought of as being sealed off from neural events in the same way that the behavior of a clock is sealed off from the laws of quantum mechanics, or the biology of cells is sealed off from the laws of quarks.

Studying neurons again makes the problem too complex. That is not to say the neurological study of the brain is useless. We can learn much about human intelligence if we could

... step back... towards a higher, more chunked view. From this vantage point, we hope we will be able to perceive chunks of program [or groups of neurons] which make each program [or group] seem rationally planned out on a global, rather than a local, scale - that is, chunks which fit together in a way that allows one to perceive the goals of the programmer [or the brain]... There is some sort of abstract "conceptual skeleton" which must be lifted out of low levels before you can carry out a meaningful comparison of... two animals [or intelligences].

That was Hofstadter again. The assumption is that intelligence can be abstracted out from the neural structure and implemented on a computer. There is of course a chance that this assumption is unjustified, in which case strong HAI is impossible.

But taking the assumption for now, what "mechanism" does a child brain (from here I will use the words "brain" and "mind" interchangeably, to match Turing's wording) consist of? In a letter from Christopher Strachey to Turing, he wrote:

I am convinced that the crux of the problem of learning is recognizing relationships and being able to use them... there are, I think, three main stages in learning from a teacher. The first is the exhibition of a few special cases of the rule to be learned. The second is the process of generalization - ie. the underlining of the important features that these cases have in common. The third is that of verifying the rule in further special cases and asking questions about it.

I think what Strachey said has credit, and I will return to it momentarily. There are, I think, other considerations. Without saying, a child brain in the form of a computer and a child brain in the form of a human has significant differences. The biggest one is the lack of physical environment, or to give it another name, reality. A lot of what the child first learns relates to the external environment. Infants at around 10 months of age learn that objects still exist when out of sight; before that, however, when infants see a toy hidden under a blanket, they wouldn't know to look under the blanket for it. This is called object permanence, and is a perfect example of how the environment helps children's cognitive development. Presumably they notice that objects keep reappearing when they disappear, and eventually realized that objects do not in fact "disappear".

How can a HAI learn such a concept without the environment around it? It may turn out that getting an abstract intelligence doesn't require learning object permanence. This would be a great test of the abilities of a developing HAI though. There is another, more serious, potential consequence of existing in digital space: the HAI may never learn language. Language is partially a mapping, through the phonetic sounds we produce, between our thoughts and objects in the real world. If a computer never "encounters" a chair, it wouldn't know what a chair is. More disturbingly, a computer never encounters the three dimensional space we live in. A number of other concepts familiar to us as humans become meaningless for a disembodied HAI. As Hofstadter wrote, "thoughts must depend on representing reality in the hardware of the brain." For a HAI to have thoughts, then, there must be some reality in which it exists whether real (such that the HAI is embodied in a robot) or virtual (a digital, artificial reality which we have control over). In either case, the environment should not only be passive, but created such that the HAI could react to it. A particularly interesting idea I had was to give the HAI a lower dimension reality - so it lives on a plane, and its "vision" consists of a colored line. This simplification serves both the purpose of giving the HAI an environment, while keeping it simple enough for it to understand and for us to understand its understanding.

Assuming these external (to the HAI) issues are solved, we now turn back to the problem of what internal mechanisms a child brain must have to being the learning process. What is the starting state, the tabula rasa from which knowledge will be written? It seems to me that the ability to recognize patterns is of utmost importance. Without it we wouldn't learn object permanence, or recognize that an armchair and a stool both belong in the category of chairs. The whole idea of creating ontologies, which is what KBs are, is based on the ability to recognize patterns and classify objects. The way we learn, too, is from recognizing patterns: we classify both hand written and printed letters as part of the alphabet, and we look at past experiences for insight on how to solve unfamiliar problems.

The last point on solving unfamiliar problems actually pushes the ability behind simple recognition of patterns to using these patterns, what Strachey called "the process of generalization". I will give it another name: induction. Note that this is not the mathematical meaning of induction - the reduction of an infinite number of cases to finite base cases - but the reasoning meaning of induction. We may burn our fingers once, twice on a hot stove, and we learn to stop putting our fingers on stoves or other hot objects. There are more abstract generalizations, too. We learn the quadratic formula by applying it on many different equations, but we know that the formula doesn't only apply for these exact numbers: it applies for any equation of the same form. This is due to our generalization of the pattern we have recognized in the equations which the formula solves.

There are admirable advances in pattern recognition. Within a KB, a Structure Mapping Engine (SME) can make analogies between two domains. This is how it works. The KB contains statements about how different components of a domain relate. SME first finds a relationship which exists in both domains; the things related by this relationship are then mapped onto one another. Each of these mappings be generate deeper mappings - that is, there are more relationships that are similar in these components. These nested relationships give each domain a structure - hence the name of the engine. By comparing not the actual objects but the relationships between objects in the two domains, a "deeper" analogy is drawn.

While this is a highly successful way of recognizing patterns and a solid step towards HAI, I can see at least one extension to the KB-SME: the pattern found should be added back into the KB as an object of its own. This stems from the fact that humans do not only reason on one level, but finds patterns made up of patterns, and patterns made up of those, and so on. Induction is the step for the HAI to create an ontological KB for itself, and is necessary for it to learn anything of significance. This is of course, easier said than done: how should the pattern be represented?

This, in fact, leads to a broader question: how should anything be represented in the first place? The idea of creating a child HAI is not complicated, but it brings into question many processes which we do not yet understand about our brains. If the child brain is to be "lots of blank sheets", how are the perceptions of the HAI written on the blank sheets? The knowledge in the knowledge bases currently in use all came from humans; some programmer/knowledge worker devised an organizational scheme, and the objects are related correct to each other. For humans there is some hard-wired method of translating our sensations into objects of thought. Children don't know what any "thing" is when they are born, but eventually they have the concept of a generic "object". Not only do they have the concept, but they can reason with such a generic object without ever having seen one. Without the abstract representation of the world, an induction engine - no matter how good - will be useless.

Not that our unbuilt HAI has a good induction engine. There is a large obstacle in this area as well: how should the actions of the HAI be represented such that the HAI can change them? To take a simple case, I am capable of using deductive logic. When I first learned it, I probably made lots of mistakes, for example, affirming the antecedent. But as I learned about logical fallacies, my thinking changed as well. I not only stopped myself from making these errors, but I am able to catch myself when I do make them. The same kind of introspection is need when we try and fail to remember a salient event (say, ate breakfast with the president), and therefore know it did not happen. The method of thinking suddenly became the object of thinking. It is perhaps not co-incidental that this forms the kind of "strange loop" which lies at the heart of Hofstadter's book. Without the ability to modify its own behavior, any HAI will still only be following algorithms and incapable of truly surprising us - not to mention not really a HAI.

The opposite end of the same problem is, what set of algorithms should a HAI not be able to change? There are processes which I cannot stop myself from doing - the best example being the recognition of faces in inanimate objects. I could not change how I look at faces any more than herring gull chicks can stop pecking at yellow sticks with red spots. Similarly, the will be aspects of a HAI which is hard coded; the perception of causality being a highly possibly example. More over, it is impossible for every line of code to be modifiable by the HAI itself. The answer to this question would not only be what needs to be hard coded for the HAI to learn, but also how little can be hard coded to have the same affect.

An additional difficulty of modifiable processes is that current KBs are ill suited to hold procedural knowledge. This runs against the current consensus that procedural knowledge and declarative knowledge is kept separately. Episodic memory should be included in the knowledge base as well, since this forms the basis of induction as well as the subjects of thought. Both these types of knowledge requires that the causal and temporal relationships in current KBs be greatly expanded and specified. It would be interesting to build a microtheory on action and its possible consequences.

I have presented a few ideas here, but also raised a lot of obstacles which the path to HAIs must conquer. I would just like to close by saying that, creating a HAI may not in fact by very illuminating for human cognition. When a HAI does start learning, all the symbols it creates will be radically different from the symbols we use; we may not be able to assign meaning to those symbols at all. Of course, that doesn't diminish the allure of creating a human-level artificial intelligence at all.