Slashdot Mirror
Making Science Machine Readable
holy_calamity writes "New Scientist is reporting on a new open source tool for writing up scientific experiments for computers, not humans. Called EXPO, it avoids the many problems computers have with natural language, and can be applied to any experiment, from physics to biology. It could at last let computers do real science - looking at published results and theories for new links and directions."
And forgive me for thinking the university would be more helpful, but no, there's been a series of expos at the University of Aberystwyth, from art through VoIP.
I'd love to have found more info on the language, but my casual browsing got stopped right there.
If they'd named it something like EXPI or EXPLO at least it'd be uniquely locatable. Google might whine about the potential misspelling of Expo, but it would dutifully locate the search term as requested.
John
Human: No Computer, Do NOT launch missle now.
Computer: Parsing input ...
Computer: NOT, NOT (launch missle now)
Computer: Launch initiated ....
After which the computers deduce they were actually not created but rather evolved from a lesser society of "users". sorry had to make the joke, we all saw it coming :)
CS: It is all sink or swim...oh and did I mention there are sharks in that water?
Wow! Now all that past work on Artificial Stupidity has REAL uses.
http://www3.sympatico.ca/sarrazip/nasa.html
WTF? If you have to manually pre-parse every article that enters the system, it severely limits the rate you can enter information into the database, no?
"In a 32-bit world, you're a 2-bit user. You've got your own newsgroup, alt.total.loser." -Weird Al
Wow...getting a machine to write up your science experiments! Excellent...now all i need is to find one that can type my essays, and show its working in my maths, and I'm sorted! Is this the new era of generating scientists from everyone!
I need one to clean my clothes, sing to me in the bath, and make sure my house is warm when I come home! Hehhe! Who needs wives...we have UBER_MACHINE
>>>Scanning for I.D.I.O.T.S. >>>
>>>I.D.I.O.T.S. FOUND! >>>
I think that computers have actually been able to do real science for at least a little while already.
;) )
John Koza is a leader in field of genetic and evolutionary computation. Very much his computer's do real science. The computers analize a set of data (observation), they make a series of modifications (hypothesis), they run fitness tests against these modified versions of the data (experiment), then they begin again analizing these results (back to obeservation).
The computer clusters which John Koza has engineered have created high-pass and low-pass filters when given nothing more than a random assortment of electronic components; even while John himself knew nothing of electronics that would enable him to create such a circut himself.
Most impressively is how the computer cluster evolved a new antenna for NASA - when it was completed John was worried that the computer had made some grievious errors because the little antenna looked like a bent paper clip - but it worked!
And that's science if you ask me. Especially the antenna - the results of experiments can, and seeminly do, often go against "common sense" and give answers which are "unintuitive".
Perhaps computers will be much better with the next generation physics we're discovering. Perhaps our little numerical darlings are simply better suited to deal with the abstract, multi-dimensional world of what the universe is starting to appear to be.
(Please pardon my lay and simplified version of the scientific method - but I feel it is a valid interpretation (if overly simplified for minds such as mine
--
Music should be free
My Computer Music Tutorial Videos
It seems to me that it's designed to fit experiments into a framework which might not allow for much innovation. The truly great experiments (e.g., Michelson-Morley, Avery-McLeod-McCarty) required new experimental techniques as well as new hypotheses and tests. We should be very careful not to impose a standard which would limit such experiments (or, more to the point, the ability of the experimenters to get published) in the future.
Basically what I'd be worried about is the tendency of the tool to become the task. This is something of a problem in my field (biostats) because SAS is so ubiquitous -- often the question becomes "what can SAS tell us about this data set" rather than "what do we want to know from this data set, and what tool should we use to find out?" Fortunately other, more flexible analysis tools (particularly R, which encourages real programming rather than running a set of canned tests) are becoming more common in the field, and so this is starting to change, but it's still a problem.
It's also a problem that every techie is familiar with -- "We want to do this in $LANGUAGE on $PLATFORM," even when that particular language and platform may be an absolutely terrible choice for the task at hand.
That being said, it's certainly a potentially useful tool, and I'll be interested to see where it goes. It's just that when I read lines like "Journals could also insist that researchers submit papers in EXPO as well as written normally," I get twitchy.
The correlation between ignorance of statistics and using "correlation is not causation" as an argument is close to 1.
FTFA: "Computers are not very good with natural language"
Neither are most humans. Even if computers could understand natural language, most people still wouldn't be able to convey their ideas correctly.
...reveals that EXPO is an OWL schema. Exactly as described, it's an attempt to regularize the content of experimental design into machine readable form (XML). So any discussion of whether EXPO is a good idea or not really hinges on whether you think OWL is a good idea or not.
How do you think the machinism behind doing calculations can possibly stop the evolution of "human understanding"? And how is having a computer doing the calculations qualitatively different from having a human do them, except that the computer (if programmed correctly, of course) doesn't make mistakes?
Don't blame me; I'm never given mod points.
Without intuition, Newton would never have discovered gravity. Intuition is not a means for conducting experiments, but it is essential in order to determine what experiments to conduct.
English is easier said than done.
What is it?
EXPO is a piece of software (written in a formal language called "owl", but they didn't tell you that), which provides a formal dictionary especially for experiments. The terms in this dictionary let you describe your experiment in a formal way. That's a bit messy, but then you're supposed to use an editor to help you. An editor for this language (called "protégé")can be fund at http://protege.stanford.edu/index.html. Download it (61 Mb., or 31 Mb. without the JVM) and use it to read the EXPO document.
What's it good for (in principle)?
Once an experiment is decribed in the OWL language using this dictionary, it can be searched automatically. You could automate queries such as "list me all published 3-factor experiments that test Ohm's law". Or "give me all 2-factor experiments that deal with lung-cancer, smoking, and gender and that use tomography as a diagnostic instrument".
Now at the moment you can do that too, but you'd have to spend quite a bit of time and know quite a bit about the field to be able to do this because you won't be able to do a full-text search (thanks to the publishers of scientific journals for this). And then you'd find that not everyone uses the same terms, and then you'll find only English-language results because you wouldn't know how to spell "lung-cancer" or "2-factor experiment" in Spanish, French, German, Chinese, Japanese or whatever, but then again neither can many foreign language authors spell it in English (which doesn't ever seem to stop them from publishing however).
Such a schema (provided it's universal and standardised like the Dewey decimal system) would allow you to find your way in the fog of language. Unfortunately however, if anything we will probably see lots and lots of different standards ("standards are good ... we should all have one !") and properietary solutions with "enhancements" and "extensions" (read safeguards against portability).
What can we expect in the next 3 years?
Nothing useful, I'm afraid. In theory it's great but don't hold your breath. Any author would have to download an OWL editor, understand the editor, understand the formal language used, and then code up his/her article in OWL using the EXPO distionary, and submit it (in electronic form) along with his article. Good luck to you authors! Lets just hope no-one makes any tiny but significant mistake in describing their experiment, and that all authors take the time to learn this formal language and then use it.
If within the nect 10 years any significant amount (say more than 5% of all publications) annually will be coded in such a schema I'd be more than surprised.
Another (perhaps the most important) goal for this type of research is a bit more subtle than replacing the Hypothesis->Experiment->Analysis->Hypothesis sequence (Scientific Method) by computers. There will still be many experiments for which human insight is the best tool for deciding a possibly fruitful idea. However, humans (i.e. grad students, who often might suggest 'workhorse' as a better nominative) are not only slower at data analysis, we are severely limited in our abilities to 'see' patterns and correlations in very high dimension data. This has traditionally limited hypotheses to extensions/reworkings of the proposed process at work in a single experiment. If computers have access to both the data and a weighted list of most likely hypotheses for subsets of the entire oeuvre on a specific subject, they could run statistical classification and pattern matching algorithms to suggest new hypotheses based on immense amounts of information. Some of these may involve a large number of variables or inputs, but there are two very significant possibilities that make this research (and certainly other projects involved in similar applications) highly significant:
1) These complicated hypotheses could still be tested relatively easy by human scientists because most computer suggestion systems for new hypothesis possibilities would likely suggest a few tests that would help to support/disprove these new hypotheses.
2) Even more simplification comes from the fact that experiments may not need to be repeated nearly as much as they do now in order to make a hypothesis -- there is an incredible amount of data already gathered, and typical AI/pattern matching algorithms keep some of the data back for testing later. If the system finds a possible hypothesis on some level, experiments as to that concepts validity have essentially already been done in a virtual sense.
3) If the somewhat positivist version of current thought in physics http://www.toequest.com/, mathematics, chaos theory, complexity theory, cellular automata http://www.wolframscience.com/nksonline/toc.html, etc. is even vaguely valid, it is quite possible that, despite the complexity and dimensionality of the input data, the 'best' hypotheses developed even by purely automated means might still be simple and elegant and/or even yield insight into possible explanatory processes rather than just statistical indicators. This would be a valuable and beautiful victory for humanism and the importance of science as a truly elegant description of the world around us.
But what happens if we get to the point where all of science is automated by computer?
We get a Technological Singularity.
"I am the king of the Romans, and am superior to rules of grammar!"
-Sigismund, Holy Roman Emperor (1368-1437)
Darn right! The universe does not fall into
hard-edged classes, at least not often.
Some good classes like "protons" and
"neutron stars" exist, of course, but
concepts like "words" and "species" are
intrinsically fuzzy if you think about them
long enough.
Same with experiments. Let's take a Linguistic
example: deciding whether or not a sentence is
gramatically correct. You can do this experiment
in several ways:
1) Give the person a sentence, a library, and
some paper. Let them take as long as they want.
2) Or, we can make it more like a conversation:
read them the sentence, and put a time limit on
it. In real speech, you have about a second
to understand a sentence, so we only accept
a "yes" or "no" if it happens within a second.
3) Make it into a reaction-time experiment.
Get them to hit a yes button or a no button
and measure how long it takes.
The point is, you can do dozens of variants of any
experiment, and any ontology will lump together
some things that are different in some important
way, or (alternatively) will split apart some
experiments that have critical similarities.
Likewise for data analysis.
Personally, I feel that Linguistics has been held
back for about two decades by linguist's expectation
that everything falls into nice categories.
I'd hate for the same thing to happen to other fields.
Just think of the Dewey Decimal system: that's an
ontology, and like all ontologies, it puts the
dividing lines in the wrong place.
All hail Science Machine!
--Rob
Towards the Singularity.