OK, someone hands you the latest Intel CPU. Could you come up with the full specs of the fab plant that made it? No mater how closely you examined it? Too hard? Lets tame the challenge. This time, we'll hand you a book describing the fab. Even with perfect comprehension, your ability to build a fab would be limited by the completeness of the book. How hard would it be to build an AI that could find and arrange all of the info necessary to build a plant that could build a chip from that perfect book describing it?

The CPU is an artifact. Artifacts are unlikely to contain all of the information necessary to reconstruct the events and resources that led to their construction. A book is an artifact. But more to the point, the contents of books are artifacts in their own right. Even with perfect comprehension, the reader of a book is missing much information that would lead to an understanding of the events, objects, and contexts of its content.

Lets start with another thought experiment. Lets say you were handed a very accurate book about a particular base ball team. What would be the likelihood that a reader of said book could then write a book explaining the game of baseball and all of its rules? How about its genesis? Its cultural implications? Its effect on a particular reader's life? What baseball says about the human attraction to games and sports?

Now lets ask another type of question. You have 3 hours to learn enough about baseball to play the game. Would your time be best spent looking carefully at a picture of the baseball field, bases, gloves, cleats, balls, bats, and hats? How about a score book containing stats on every baseball game ever played? Wouldn't you learn more and faster by watching a game being played? Our brains are especially good at laying down memories such that understandings can be derived at levels that match the actual subsystems interacting to produce the original artifact. Or is it that systems interact in predictably ways because the underlying causal laws of physics dictate layered ontologies? Either way, it would seem obvious that events produce artifacts and not the other way around.

Despite this, most of the AI work of the past ten years has had to do with document parsing… gleaning meaning from pre-existing media. I appreciate the boldness of this challenge, but its a lot like learning to box by fighting a heavy weight title battle with Muhammad Ali. No?

If someone asked you to learn Serbian, would you be inclined to go to the Serbian national library and work word for word through every book? I doubt it.

We have all eaten thousands of meals, does that make us good cooks?

Compare the daunting task of post-assesement (understanding cooking by picking through a meal) with the structured linearity of real time observation (watching someone prepare that same meal). The difference should be obvious. The grammatical layering that is hidden in a plate of prepared food is laid out sequentially to anyone observing the actual process as it unfolds.

Let's imagine an AI researcher. Let's assume that this AI researcher has limited resources. Let's say that that this AI researcher wants to build a system that knows things. Lets say that our AI researcher understands that it would be even better if the system knew how to learn. Lets go one further and assume that our AI researcher wanted the system to learn how to learn, for itself.

In its capacity to learn, such a system would be like a child. In its ability to self acquire the ability to learn, it would mirror the process of evolution. If you wanted to minimize the time required for a child to learn or a system to evolve, it would seem obvious that you would want it to observe and to build and to understand causality as a result of watching systems precipitate from other systems. No?

Imagine the absurdity of the opposite situation. Imagine a universe that must understand itself before it builds itself. Obviously, understanding, the ability to acquire an abstracted, navigable map of reality, requires a greater quantity of complexity handling capacity than is required of the subsystems, the materials and forces, of which it is built.

Yet is precisely this re-ordering of the causal stack that top-down parsing schemes demand.

It is impressive when stochastic pattern matching tricks can yield useful "answers" to questions about "knowledge". Proximity indexing schemes let loose on large enough sample corpus can and do produce output that when fed as source to a large enough language processing system (our brains) will suffice to trigger non-random association patterns (learning). But by them selves, such schemes are poor substitutes for the layered grammars that are acquired in learning that results from more primary and intimate knowledge acquisition that occurs through direct interaction with the causal systems of which text is merely an abstraction.

Corpus apologists rebut such logic by pointing to the linearity of text. They say that this linearity preserves the context of direct process and observation. But text is simply a triggering mechanism. Both writing and reading are resource expensive activities. We write only that which is necessary to set off the cognitive cascade that we experience as knowledge. It would be foolish to assume that any such text would ever cary the full semantic content engaged during reading (or writing).

Revisiting the question I posed at the start of this post, wouldn't it be far easier to build an understudying of chip fabrication (and chips for that matter) if one had access to the full historical record (in temporal sequence) of all of the actions taken by all of the entities involved in its design and construction? Sequence tends to reflect aspects of the physical relationships at play in any system. Ontogeny recapitulates phylogeny!

When AI researchers write top-down parsing schemes, they start with raw text and work it into tokens. Tokens are standardized and simplified statement forms "subject, verbs object". The tokens are then listed in indices that contain references to each instance in which they appear and pointers to other indices that associate synonyms and related terms. Tricky programmers feed proximity data into these abstracted graphs such that stochastic values weight the likelihood that a term will reference a particular meaning depending on surrounding terms. Google and others build in authority weighting that assigns greater value to terms referenced from other authors.

It is highly likely that our own brain uses similar techniques. But brains aren't shaped solely by static texts. Lucky for us, brains are shaped first by years of direct observation of events and interactions.

My contention is that the indexing schemes currently in use are more than sufficient. Deficient is the source data being fed to them. What's missing is a training phase in which the data being fed into the graphing scheme is primarily observational.

If a project existed to extract observational data from the daily sequence of the stuff people do on their computer or smart phone (think, SETI@home, only you are the alien) and it required that you install a chunk of code that both collected this data and uploaded it to the giant parsing machine in the cloud… would you?

Randall Reetz