[Search for users]
[Overall Top Noters]
[List of all Conferences]
[Download this site]
| Title: | Europe-Swas-Artificial-Intelligence |
|
| Moderator: | HERON::BUCHANAN |
|
| Created: | Fri Jun 03 1988 |
| Last Modified: | Thu Aug 04 1994 |
| Last Successful Update: | Fri Jun 06 1997 |
| Number of topics: | 442 |
| Total number of notes: | 1429 |
248.0. "AI is growing up/changing" by ISTG::SELECT::KELLY () Thu Nov 08 1990 18:32
Greetings,
The AI Forum report in the note above points to an
interesting change in the AI universe that we've also seen here.
AI seems to be growing up and assuming its rightful place in the business
world. It seems to be becoming what we've always been telling people it
was: just another tool for solving business problems, not magic.
In the informal discussions here among KEs, we've been
trying to anticipate the kind of work we'll be doing five years from
now. The one conclusion we've reached is that we won't be doing the
same kind of work we're currently doing. But there are a gang of
different opinions on what we'll be up to in 1995.
Some of the factors we're considering are:
- We get customers now who not only view AI as NOT_STATE_OF_THE_ART,
but as actually DEAD. They claim that neural nets will take over all
the tasks that knowledge-based systems have done till now.
On the one hand it's easy to dismiss these guys, because they are the
same knuckleheads who predicted AI would eliminate Data Processing,
Scientific Programming, and Operations Research/Linear Programming.
And most of them couldn't tell a neural net from a rectal sphincter.
But they do have some good points. For example, it took about three years
to build the excellent DECTALK product by long and laborious
encoding of pronounciation rules. But a neural net recently
built (NETtalk) was able to reach the same level of proficiency by
being trained overnight (took 6 months to build the net tool).
So, there are some tasks that we've been trying to do with
knowledge-based systems that are better done with nets.
- The Gulf Crisis now looks utterly dismal, and most Americans are
mentally preparing themselves for a war in the middle east (started by
Bush). If the war comes, energy costs will skyrocket, and business in
the US, Europe, and Japan will slow to a crawl. Companies don't
purchase computers when they're suffering. They wait for a few years.
This will mean that even our biggest customers now will delay hardware,
software, and services purchases. And for AI it could be disastrous
because most companies still put AI in the R&D category, and R&D is the
first budget to be cut in a financial crisis. So, AI consulting
contracts may just completely dry up.
- The technology has progressed. Some people seem to have come into AI
because it was new and exciting. They liked the glamour rather than the
hard work involved, and those people are moving on to other things.
In a way, this is a huge relief because we've been trying
to convince people for years that AI is not magic. But it's a serious
career crisis for people who have bet the whole farm on knowledge-based
systems - people at DEC and elsewhere who have spent the last ten years
carving out a niche for themselves as AI managers in the company, only to
see the field splinter up into several fields or become prosaic and
unexciting to customers. Obviously (we can see this happening here)
these folks are now trying to either convince people that nets are silly
and useless academic fluff, or are trying to call themselves
K-based AND Neural Net kings - hoping to gain more power, rather
than losing power to people doing neural nets work outside their groups.
Fascinating politics involved.
- Five years ago, most of the companies we dealt with were just
starting their own AI groups and they wanted us to train them and
co-develop expert systems till they got going. Now, virtually all large
companies and many small companies have an AI department. So, we're
just being called in to assist on a project, or to take a part of a
larger project, or to work alongside the company's AI group on a
difficult project requiring more KEs than the company has. We used to
be ground-breakers. Now we're more like contract programmers.
- The AITC has been undergoing significant changes. The number of
people working on internal projects and product development is
shrinking and the number of people doing revenue generating work is
growing.
So, we have all these things in mind when contemplating the future.
AI is pupating. It's changing its shape, and we figure we'll have to
adapt or die.
What we're doing here, as a result, by a kind of tacit consensus, is:
- We're expanding our horizons and we're generalizing. Most KEs here are
teaching themselves Nets. We'll teach our first DEC course on Nets
(offered by myself and Ram Josyula) in January. The first course
(one week) will be only for DEC KEs. Then we'll take the show on the
road to customers. We've been turning down net business till now, but
we'll be able to capitalize on it next quarter. We're also pursuing
Artificial Life, induction/case-based, genetic algorithms, anything that
we think will be of use on projects. Not having enough tools to work with
was always a complaint of KEs, but now we have too many tools - and
the tools are in fields that require radical re-education, not just
the learning of a new language. So, we have to jump in head first, stop
moaning about the changes AI is undergoing, and start using those
changes to our own advantage.
- We're not snubbing our noses at integration work. Expert systems
have to be linked into a company's existing computational infrastructure.
We've always known that, but we've tried to do "just AI, pure AI".
No longer. AI has to be useful to succeed and integration makes it
useful. So, we do it.
- We admit that AI no longer has any glamour. But we also admit that we
were never selling the glamour; we were selling a useful tool. And we
still are.
- We make sure we have enough skills useful in any field to get us
through the economic depression that some people are predicting.
- And we learn everything we can to improve our skills. AI is changing.
The world is changing. And DEC is changing. During the good
times, when the business environment was stable, the specialists
flourished, But when the environment changes, as it is doing now, the
specialists die and the generalists survive. So, we keep learning.
Dikk
| T.R | Title | User | Personal
Name | Date | Lines |
|---|
| 248.1 | Using AI's changes to our advantage | ISTG::SELECT::KELLY | | Thu Nov 08 1990 18:43 | 118 |
| To illustrate our commitment to change, here are some of the Nets
projects we've been building with external customers in the US
Fellowship Program.
Protein Folding:
The most recent app is in pharmaceutical manufacture. The
problem has to do with the stoichiometric configuration of synthesized
proteins. We have a linear decryption of a particular enzyme, for
example. (We're starting with Bovine Phospholipase A2 which, beside
being found in the livers of cattle, is found in snake venom and in
the lungs of human asthmatics - hence the customer's interest in
synthesizing the protein.) But we don't have the ANGLES at which the
segments of the compound are joined. That is, we know the sequence
(carbon-carbon-phophate-benzene ring-....) but we don't know how the
elements in the sequence are twisted around each other (does the
benzene ring jut out at a 45 degree angle?). This kind of stereochemistry
problem (guessing that there is an alpha helix here or there)
results in repeated testing (11,000 compounds tested to get one new
useful drug). So, we're starting to use the net now as a complexity
reduction mechanism to find likely chemical cousins for a particular
linear sequence. The training set contains the database of ribbon
tracings (a way of representing molecular structures) mapped to their
respective linear sequences. This should give us matches on similar
structures. But we're only beginning here. We hope to also use
structure fragments as training elements, and eventually synthesis
data as well - allowing us to propose a method of synthesizing the
compound.
Flood control:
Rain clouds bounce back a certain signature pattern when they are
hit with radar radio waves. This signature data was correlated with
actual rainfall measurements beneath the clouds. In this way it was
possible to get elements for the training set: a complex radar signal
(numeric, rather than graphical), and a reading of the amount of rain
beneath the clouds being radared. Those radar pattern-rain guage data
matches trained the net. And new radar data could be matched against
the learned patterns and rain amounts identified/interpolated.
Image recognition:
Realtime video images (still frames) of underwater structures are
beamed up from a mobile submarine searching for structural defects in oil
derricks and other underwater structures. Those pictures are pre-processed
(digitized and smoothed) and handed to a net which recognizes where on
the underwater structure the submarine's camera is pointing - based on
comparison to a training set/database of oil derrick structure fragments.
This allows the submarine to learn its position vis a vis the underwater
structure without requiring steerage from above. A transformation
algorithm was applied to compensate for trajectory and perspective.
Steelmaking:
A tundish gives off a recognizable complex pattern when it is
emptied or refilled (a tundish holds molten metal and pours it for the
formation of ingots). This pattern can signal disaster or a normal
pour. The patterns are roughly sigmoid, so the task is a kind of
signal processing. And the realtime needs of the system are relaxed
(realtime in this case means minutes for interpretation, not
nanoseconds). So, we don't have to worry about the monstrous
cpu-intensive dragging that might kill other net applications in a
realtime environment.
Financial:
Technical market analysis is obviously an ideal use of nets
because elfin pattern recognition (head and shoulders, wedges, penants,
inverted yield curves for bond trading, etc.) is based on both experience
and pattern recognition/feature extraction abilities. But the object
segration and transformation problems are turning out to be a bear. The
trouble is that a pattern (an input to the training set) may be comprised
of six input elements or sixty or six zillion - any of which would signify
the same pattern. For example, head and shoulders is up, down same, up
twice as much, down same, up half as much, down same. But that pattern may
be mapped in six data points (six days trading) or in ninety-three or any
number. That makes for a data/pattern condensation problem. The other
nasty bugger is the object segregation puzzle. Most of what appears on a
techician's charts is slop. But the diamonds can be extracted from the mud
by a discerning eye. Training the net to only look at the appropriate spot
- to become the discerning eye - seems to require building a net within a
net (using feature extraction as a kind of object segregation); or sliding
a pattern over a graph while hoping to strike a match. We're still
puzzling out this one. Still makes my brain hurt. But we'll get it.
We have no option. We have to get it.
Disposition of Projects:
I'm hesitant to mention customer names (all the work is done under
non-disclosure), but I'll give general data, and folks can see demos
if they're interested.
The Submarine project was done with an Italian energy
exploration/exploitation company. The tool was PAM. The company has since
bought PAM and you may see the project demo sometime in the near future
at a European or US AI or Nets Conference, maybe IJCAI.
You can see a demo here. Contact Ram Josyula (SELECT::, or ESIS::).
The radar meteorology system was done with another Italian company.
The tool was CASCADE - which is public domain. The company
is expanding the original system using a simulator to spit out new
rain guage data. To see the demo, contact Dave Cavallo (SELECT:: or
ESIS::)
The tundish project is for a Korean steel manufacturer, and the
financial project is for a Korean wire house.
We're still testing tools. These projects are not yet done.
In about six months we may have demos ready. We'll see.
The pharmaceutical project is with an American drug-maker. This is
a hot topic in the industry, and if we cracked the whole ugly
problem we'd be assured of a Nobel Prize, so we only hope to
peel off an edible piece. This one is in the early building stage.
But check back in four months. Whatever part of it we do will be done
in four months.
Dikk
|
| 248.2 | Thanks Dikk | HERON::ROACH | TANSTAAFL ! | Fri Nov 09 1990 14:25 | 10 |
| Thanks Dikk - Super stuff as always - I believe that we all need to do
more than just manage the change that is impacting us; WE NEED TO
EXPLOIT IT!
I'm happy to see that two of the NEW approaches were on European
projects. I have checked, and we can talk about the submarine vision
project. It was done as a Fellowship project with the Italian
equivalent of the US Department of Energy, ENEA. I understand that the
fellow has just delivered a paper on the project at a convention in
California.
|