New York Times on the Web Forums Science

Missile Defense

Russian military leaders have expressed concern about US plans for a national missile defense system. Will defense technology be limited by possibilities for a strategic imbalance? Is this just SDI all over again?

rshowalter - 12:19pm Jun 27, 2001 EST (#6133 of 6168)
Robert Showalter showalte@macc.wisc.edu

In MD6132 rshowalter 6/27/01 10:45am ... there's this:

" The main stumper has happened because the AI community funded by the military came up against a mathematical constraint that it was clearly warned against, ignored the warning -- and has been trying to make big progress along a line of work where big progress is provably impossible -- for a decade.

" If checking had been morally forcing within that community -- the US would have better weapons today -- and a less frustrated and corrupted cadre of classified researchers, as well.

The artificial intelligence efforts funded by the military have been dominated by a "connectionist" or "parallel distributed processing" paradigm that was showing severe limits by the late 1980's -- when really serious efforts to make hugely parallel chips were built by the military, to try to get connectionist systems fast enough for really good missile guidance. Funding in the open literature continues to be dominated by these connectionist models - indicating the classified people are still working with them. The problem with these networks, for almost all the practical cases -- is that the get slower (in number of cycles per calculation) VERY fast as the complexity of the networks increases. So that a 10 times bigger network can easily be billions of times slower (in computation cycles). This is something like a "brick wall" -- where very large increases in computation power yeild only small increases in performance.

rshowalter - 12:20pm Jun 27, 2001 EST (#6134 of 6168)
Robert Showalter showalte@macc.wisc.edu

In 1990, J.S. Judd , a young pre-tenure acacemic, wrote NEURAL NETWORK DESIGN AND THE COMPLEXITY OF LEARNING MIT Press, 1990.

Here is Judd:

“ . . . . “The published successes in connectionist learning have been empirical results for very small networks, typically much less than 100 nodes. To fully exploit the expressive power of networks, they need to be scaled up to much bigger sizes, but it is widely acknowledged that as the networks get larger and deeper, the amount of time required for them to load the training data grows prohibitively. . . . "

Judd means something compelling when he uses the word "prohibitive." He shows, by mathematically accepted standards, that even "simple" rote learning, at the scales animals do it, is impossibly slow (the technical term is NP complete, which is taken as the standard demonstration of intractability in the computer and crypto sciences. )

No one called the result wrong, but people found ways to ignore it.

The response of the neural modeling community, with funding and work patterns dominated by the military, was to ignore the result, and marginalize Judd.

Judd was denied tenure, after writing what I believe was an outstanding book.

Events since 1990 have tended to show that Judd was right. Progress in "connectionist" neural modeling has been, if anything, slower than Judd's results might have predicted.

For the reasons Judd was clear about in 1990, progress in the artificial intelligence that the military cares so much about has been VERY slow in the last decade.

There are plenty of examples where "the digital revolution" makes enormous progress possible. Convenience of calibration, for systems where the physics is fundamentally stable within the calibration range, is an example. There's no reason to dispute that - it should be celebrated.

But, for mathematical reasons, there are many cases where an explosion of computational power gets much less than one might expect. There are a number of such examples that one would expect to occur, judging from what is known in the open literature, in the control problems of missile defense.

rshowalter - 12:25pm Jun 27, 2001 EST (#6135 of 6168)
Robert Showalter showalte@macc.wisc.edu

On war -- if it is all right to kill anyone associated with a name such as "communist" -- then one can justify anything at all.

At a fundamental level, much of the mass death in Vietnam caused by American military action does not look any better, morally, than much of the mass death produced by the Nazis.

If you think otherwise, you can pretty quickly get to stances that "make Machiavelli seem like one of the Sisters of Mercy."

Nazi war criminals often argued that their pattern of killing was better, not worse, than bombing -- and it is hard, from my distance, to see exactly what is wrong with those arguments.

In Korea, just to take an example, American fire and dam bombing killed two million people, mostly civilian. Was this somehow purer than what the Nazis did?

rshowalter - 12:38pm Jun 27, 2001 EST (#6136 of 6168)
Robert Showalter showalte@macc.wisc.edu

This approach does not have the same limitations connectionism has -- in many cases, it can be billions or trillions of times faster doing jobs control systems need -- I put in on the internet, and Kline informed people about it, in the early 1990's -- as I'd been told to do - - and waited to be contacted -- as I'd been told to do.

It talks of "neurons" when it should use the term "glia" -- but the math is simple, and I tried to block it out clearly.

The approach is fast in digital form -- but can be much faster if some of the components are calibrated analog - using technology that has been mostly available for years.

http://www.wisc.edu/rshowalt/pap2

gisterme - 12:54pm Jun 27, 2001 EST (#6137 of 6168)

"...GI: Putin link actually started here from London: MD5751 lunarchick 6/22/01 9:20am ..."

Thank YOU then, possumdag. Great link, wherever it came from.

gisterme - 12:55pm Jun 27, 2001 EST (#6138 of 6168)

...or lunarchick or wherever it came from. :-)

(30 following messages)

New York Times on the Web Forums Science Missile Defense