Legal
The views expressed in this blog are my own and do not necessarily reflect the views of Adobe Systems Incorporated.
Search
November 15, 2005
Old versus New AI
I just returned from the International Semantic Web conference in Galway. A very good conference delivering a large number of important papers and convincing me that semantic Web applications are beginning to emerge into the real world..
But later I was discussing AI techniques with some folks at PARC and at Adobe the other day and a couple of questions arose.
My comment was that while the current crop of AI techniques work and work well, they do not reflect what we do when we reason or when we perform instinctual activities. The question came back, "Does a computer have to do things the way we do to be intelligent."
My answer was "no, not at all". You cannot argue with the success of fuzzy logic, vision processing algorithms that are increasingly able to understand the contents and purpose of photos, document understanding systems that can restate the menaning of text in new words. These are all amazingly powereful and successful technologies.
However, my other comment was that it would be fascinating to work on the other problem, to reproduce what we do when we think, plan, understand, and create And I do not believe we will achieve emergent intelligence similar to our own until we pursue that harder and less travelled path.
IBM and EPFL are pursuing this to some degree by modelling the neocortex. Info is at:
http://domino.research.ibm.com/comm/pr.nsf/pages/rsc.bluegene_cognitive.html
This is an ambitious project but one worthy of IBM's Blue Gene super computer.
The other question that arose was "why is it that when we build AI systems, the ones that work are no longer considered AI?" This has plagued the AI arena for years. it seems that as soon as we develop a new technique for deriving a semantic result such as decomposing unstrutured text into structured document leements using heuristics, controlling temperatures in laundry machines via fuzzy logic, or recognizing ancient script from illuminated manuscripts, it is immediately classed as useful algorithms, seantic technology, vision processing, anything buy AI. Yet, even when we start by saying we will research an AI problem, the solved ones must move into another descriptive space.
I suspect some of that is because we do NOT solve these problems with emergent AI, intelligence that both arises out of some "ghostly signature" in our knowledge bases and tta seeks to extend itself. It is, i contend, the very absence of these traits that makes us so reluctant to embrace wht we HAVE accomplished as Intelligence.
In some region of ourselves, we know that what these techiques represent, as amazingly successful as htey are, is Artificial Smarts, not Artificial Intelligence. We sense that Intelligence will be recognizable and will be emergent, not algorithmic in nature. And we won't be happy with AI until we get that.
Comments?
Bill
Comments
The paper "Architectures for intelligent systems" by J.F.Sowa mentions this. He says that parsers, search-engines, game playing programs are all AI components. He further says: "Their theoretical foundations are much better understood, and they have found their way into applications that are no longer considered part of AI."
I had written a blog sometime back about whether "Semantic web" can be considered AI or not. Although, W3 says it is not, the question ultimately comes to what AI is all about?
All the while I thought you were talking about Adobe Illustrator, then it hit me- she's talking about Artificial Intelligence! How come you're here. Must be these guys manning the Adobe blogs are not reading the blogs at all? Wow
---------------------------
Yang,
Sorry to confuse you with the AI term. Yes, I am specifically discussing artififial Intelligence, not any product.
We do read the blog entries of our colleagues, but We're running out of acronyms in the industry :)
Thanks
Bill McDaniel
"In some region of ourselves, we know that what these techiques represent, as amazingly successful as htey are, is Artificial Smarts, not Artificial Intelligence. We sense that Intelligence will be recognizable and will be emergent, not algorithmic in nature."
For the development of, what is stated here as Artificial Intelligence, and not Artificial Smart, "Emergent" is indeed the keyword. Which, in my opinion, most fundamentally assumes the concept of freedom.
I am no expert in the field of AI, and I've just recently read a few things about Neural Networks and AI applications and theory in general, but my belief is that we are asking so much but giving so little. In other words, research aims so much to practical outcome. Utilities. Can these two notions-Utility and Intelligence- coexist, in a philosophical sense? Let's look the other way round. What if we trained a human, in nothing but, say, character recognition. No other stimuli except a bunch of character-sound associations. For years. What would happen? Could this man, being raised in a severely controlled environment, show any kind of intelligence? Chances are that, he couldn't. Freedom and complexity, as I see the matter, are definite to the emergence of any kind of what we call intelligence (although I can't grasp the term as tightly as I would like to).
Regarding complexity, I am sure that at present we do not have the means to create. But what about freedom?
Excuse me for my somehow-poetic writing, but that's the best way I could cram it up, considering its almost 3 in the morning;)
----------------------------
John,
Thanks for taking the tme to reply. I agree with you. The problem with much of our attempt to investigate and explore intelligence by artificial modelling (AI by another name) is that we want results, utility, and functionality.
The techniques in place have yielded that. we have some very smart systems in place or coming soon.
But we do not have much work or time or money being expended on emergent intelligence as a function, peraps, of complexity. We need more.
With respect to your concept of raising someone only to recognize characters. First, let me say that the lack of stimulus for such a person would play havoc with development. I speak from personal experience. One of my nephews is an adopted child who was allowed by his natural parents to lie in a crib with a bottle and nothing more for stimulation for days on end. The rsult was a child with no language, severely delayed cognition, little emotional control, hearing, teeth and heart problems (my sister and her husbands are saints...they adopted him at 5 and have brought him forward to 20 with love, education, and discipline. Some of the damage can not be undone, but he's a great guy now).
An emergent intelligence will, I am sure, depend not only on the complexity of its underlying structure, but also on the type and variety of its input, its stimulation if you will by the real world.
Thanks for commenting...all poetry accepted...we may not have the objective langugae to discuss this yet.
Bill
I think your thought on the gap between what we believe to be "real AI" and the useful algorithms formerly called AI is mainly due to a deep sense that we have missed something.
The "emergence" property is one way to articulate it. In the end, however, the bottom line is that it seems current AI is always emulating rather than abducting. It is the failure to generate flexible, effecient, and executable representations that has held us back. Real AI would be able to program itself once it had bootstrapped from data. We sense this intrinsically and, thankfully, we have the humily to know that we have not yet arrived where we need to go.
There are some people who are working and showing success in getting to the above kind of solution one interesting problem domain at a time while maintaining a generalized approach that is not domain-specific.
--------------------------------------
Tim
I like your comment that current research and development is emulating not adbucting. I think that states the issue very well. It is good that we realize that what we do internally is not what we emulate in our machines. I sum this up with my comment about not solving differential equations when I catch a baseball...or view a Monet.
I would be interested to know of anyone who is specifically working on uncovering that emergent process rather than just more blunt calculation tchniques...which, as I said, work wonderfully, just not like me.
I think one of the things we need to look for are patterns in knowledge representations that are sort of lurking beneath the surface. Representing the same knowledge multiple ways and then looking for hidden representations that cross the boundaries might be a way. i often refer to seeking the Phase Space of KR. That representation that abstracts the factual away from the underlying pattern.
Thanks for writing and I'm looking forward to talking with you soon
Bill
The AI Effect is the name of the gap that you mention.
--------------------------------------------------------------
Thanks
Bill