[12]                                home                            [14]

ORB Visualization

(soon)

 

 

Colloquy between John `Josephson and Paul Prueitt

 

switch to Technology Innovation beads -> {<%>}

 

 

 

 

Paul’s comments are indented, second response are indicates as {1}, {2}, {3},{4}

John’s responses are not indented

 

> A great deal of value is obtained by finding functional

> decompositions of inferential actions by humans, what is called

> cognitive engineering at DARPA. 

 

 

> Our problem is that the Nation needs a breakthrough that cannot occur

> as long as the AI entrenchment remains.  As you so properly put it,

> “block heads in high places” reinforces the entrenchment.  

 

This "AI entrenchment" is a complicated business. For the record, I am employed by the Computer Science and Engineering Department of a major university as a faculty member with a specialty in AI, and associated with the AI laboratory. In this sense, at least, I am an "AI researcher."

 

> Our position is that inference is an entailment, like

> physical cause. 

 

For cognitive agents, inferences are part of the causality that determines behavior?  If that is what you suggest, then I agree. {1} More generally, for any information processors, insofar as they are information processors, the contents of the information (whether this or that is coded) determines their behavior (at least partially). "Inference" is a kind, or family of kinds, of information processing, involving manipulations of representations of the world or environment of the information processor.

 

> Human inference is only poorly reflected in classical logic,

 

Agreed.

 

> and the full nature of induction is barely touched by formal thought

 

In my judgment, probability theory has made profound contributions, but it's not enough.

 

>  – in spite of its key role in axiomatics and theory proving.

 

I agree that theory formation in science and everyday life is "inductive" in the broad sense of the term.

 

> This position is inconsistent with 95% of the Artificial Intelligence

> curriculum taught in America’s colleges.  The issue is simply ignored

> in the AI textbooks.

 

The position being that classical formal theories of inference are quite insufficient to characterize biological or human intelligence?

 

I suppose this is controversial, but I don't suppose there will be much support for the sufficiency of classical formalisms for characterizing intelligence among those who have thought about it.  Maybe AI textbooks should make more of this point, but they have to teach what can be taught, and relevant parts of classical formalisms are a reasonable place to start. Textbooks aren't supposed to be too controversial.  AI as a discipline has been heavily influenced by psychology and other cognitive sciences in various ways, and at various times. {2}

 

> The work in AI is made with great dedication and is elegant.  But AI

> is taught in computer science department as a discipline to be learned

> as opposed to a natural science where principled investigation of

> natural reality grounds the future work of the student. 

 

It is probably correct that AI, as taught, commonly pays too little attention to biology, and to "intelligence" as a natural phenomenon. This probably varies with the school and teacher.

 

I approach AI as part of theoretical biology, where we are interested, not just in intelligence as it occurs in humans, or more generally in terrestrial biology, but even more generally, as the study of the principles by which intelligence can function, whether in humans, extra-terrestrials, or computers.  On this view, AI studies the effectiveness of possible designs for intelligent information processors. {3} Methodologically, there is a test of rigor in theory forming, which is implementability.  Theories have to be at least implementable, or better, implemented and tested.

 

I agree that AI is too often narrowly thought of as merely an engineering discipline, or merely a mathematical discipline within computer science.

 

> As a consequence, the discipline of AI has been populated with tens of

> thousands of individuals who will not look at the concern from the

> natural sciences, and who are rewarded by this blindness from the

> federal funding sources and from the tenure system.

 

Federal funding sources and the tenure system are blind to the importance of biology to the discipline of AI?  Federal funding sources are usually mostly interested in applied AI, that is, AI as engineering. And very few engineers know much about biology.  Many "AI engineers" don't care much about AI theory, and think of AI as a bag of techniques loosely inspired by observations of people.

 

> The fundamental blindness of AI researchers

 

I plead not guilty to ignoring biology. {2}

 

> to the expressed concerns of the natural sciences

 

What concerns, exactly? {2}

 

> has become a greater and greater problem over time, while federal

> funding continue to cause a type of attention deficit disorder in this

> regards. 

 

Federal funding draws attention away from what, exactly? {2}

 

> ...

> The incorrectness of this work is not a personal attack on the

> dedicated efforts of thousands of PhDs in AI.  But the incorrectness

> can be easily characterized. The computer processor does not

> “experience” the state transitions in silicon,

 

Are you sure of that? :-)  Do humans experience the state transitions in neurons?   {2}, {3} Surely there is more to computation than state transitions in silicon, just as there is more to human information processing than state transitions in neurons. I don't know if digital computers do, or can, have vivid subjective experiences or "qualia",  but I am not convinced by the arguments of Searle and other skeptics that digital computers cannot.  It is plausible to me that "consciousness" is not reducible to information processing in much the same way as life is not reducible to carbon chemistry.  It is "implemented on" chemistry, to be sure, but higher levels of organization and structure are where most of the action occurs. { ? }

 

> it does not “know” the information that is constructed or moved. 

 

I think that "knowledge" is a functional property of systems, and computers can have knowledge, if they process representations to achieve functionalities related to effective action.  { ? } We might say: "Knowledge is true representations that are used effectively and their truth is indispensable to explaining why their use is effective."

 

> ...

 

 

> Deductive inferences and deductive chains are abstractions that can

> and will diverge strongly from natural reality in complex systems like

> social and psychological systems. 

 

Agreed, but probably most AI researchers work outside the paradigm of deductive chains, though I think many in AI and philosophy (e.g., Fodor) are confused about this.

 

 

> This was the core issue presented over and over at our three day

> conference at NIST.  Many of the participants spoke to this core

> issue, while the two ladies from the AI division at the National

> Science Foundation where so kind as to correct a natural definition of

> what inference is, by scolding the notion that inference was anything

> other than as AI formally defines it to be. 

 

I thought their point was that "inference" should not be confused with "deductive inference," and most researchers in AI are not confused this way.

 

 

> Thus the work on understanding knowledge and perception from an

> inferential and computational perspective is likely to not find

> grounding in the natural sciences, in spite of the massive US federal

> funding that has occurred in support of this discipline.   

>

> This support is being called into question in a proposal to Establish

> the Knowledge Sciences.

 

You propose to de-fund AI in favor of "Memetics"?  { No }  Or de-fund some forms of applied AI in favor of a view of Knowledge Sciences outside of academic Cognitive Science, and conceived as being better grounded in the natural sciences? {2}

 

 

> Like the Manhattan Project, the National Educational Project to

> Establish the Knowledge Sciences is designed to change the nature of

> the social discourse.  It is said that the Manhattan Project changed

> the nature of war.  The National Educational Project could change the

> nature of information technology and mass media; by bringing a deep

> and natural understanding of computer science and mathematics in line

> with the natural sciences. 

 

If the main argument is based on bashing AI, then count me out! {4}

 

> .... But one might wish to look for a first principle that is not

> captured by formal constructions,

 

A principle that is beyond mathematical representation?

 

 

> something like inductive processes that are not reducible to deduction

 

Or, rather, processes where reduction to deduction obscures how and why they work so well.

 

> but which rely on direct measurement of reality as it is in a present

> moment. 

 

As in robot perception?

 

> Implicit in your note is the statement that a departure from the AI

> myth

 

What is the AI myth?  If it's that digital computers are capable of intelligence, if only we know how to program them, then I do not at all agree that this is a myth.  It certainly cannot be shown that something extra is required beyond having the right kind of programs.

 

> or from softer forms of the AI myth, such as cognitive engineering as

> defined by funding sources, will turn the funding off. 

 

Right now my funding is based on the promise of solving actual, practical problems, at which I think I am making good progress.  My funding would, of course, be negatively impacted if is were widely believed that knowledge systems, of the sort I think I am building, cannot be built.

 

> This phenomenon is to be the subject of a Congressional inquiry. 

> (being considered as of 3/19/2004)

 

Whether "strong AI" is possible?  So congress is going to have hearings on the metaphysics of the mind-body problem? Or whether engineering AI is too attached to deduction?  So congress is going to have hearings on whether the appropriate logical analysis is being used in most of technical AI?

 

> We strongly feel that the exclusive, and highly funded, AI paradigm

> reinforces the problem of abduction to false premise.  If the

> computational inferences, which must be deductive, are composed into

> complex chains without human-centric inductive inferences, then we

> have a foundational problem leading to false sense making. 

 

I agree that sense making critically depends on abductive inferencing, and that abductive inferencing is poorly represented classically.

 

> Abduction, defined as you have as a computational inference to best

> solution

 

I would say, "... to the best explanation."

 

> can be used to support unwarranted conclusions. 

 

Abductions are, of course, fallible, but I claim that they provide warrant for their conclusions.

 

> As discussed by others at the Friends of Intelligence Community

> meeting, a premature closure can and is often occurring in real

> everyday intelligence analysis. 

 

These are failures of abduction.  They fail to sufficiently consider alternative explanations.  You can think of premature closure as a "fallacy of abduction" rather than an inadequacy of abductive inference, per se.

 

> The literature on computation and information and computational

> inductive inference is extensive and has been highly supported since

> the early 1950s.  The alterative paradigms have not been supported. 

 

I don't know if "connectionism" counts as a counterexample or not.  It certainly had a period of generous funding.

 

> Consideration is made in using Waste Fraud and Abuse laws in order to

> demand that the funding system have more transparency and be required

> to take responsibility for large expenditures that end up being

> rejected by the community of intelligence analysts as being more of

> the same poorly designed and poorly performing software.  Honesty

> about how federal dollars are spent is required. 

 

What makes you think that software intended to support intelligence analysis is rejected at a higher rate than any other new software put in front of any other users?

 

> The dollars can be spent more wisely and achieve greater social

> benefit if the artificial maintenance of the artificial intelligence

> academic community is eliminated. 

 

!  Academic AI is, of course, quite exploratory, at this stage of its short life.  The quality is uneven, and sometimes ideas are pursued more as fads than as normal science.  But, so what?  What is the positive idea being put forth as an alternative?  And why is any such alternative not simply an argument Within academic AI, as opposed to an argument Against it?

 

>  ...

 

Gotta go.

 

  .. jj