You are currently viewing SemiWiki as a guest which gives you limited access to the site. To view blog comments and experience other SemiWiki features you must be a registered member. Registration is fast, simple, and absolutely free so please, join our community today!
Ippisl - nice article, much more intelligent than some of the things I have read on the same topic (eg an editorial from Wired a few years ago). Of course any level of help and new tools should always be welcome - to reduce data, to intelligently search papers, etc, etc and perhaps even to suggest models in some cases.
The reason my support is not completely generous is that I see some weaknesses in these methods. First the approach helps purely with empirical science. There are times in scientific advance when progress is indeed made through experimentation and these methods will absolutely help there. But there are also times when advances are made by bold theoretical leaps and I don't see how analytics helps make those leaps. Maybe this is OK, but it rather undermines the claim that science can be greatly accelerated if only certain phases can be accelerated. Maybe it gets us faster from leap to leap, but we're still stuck waiting for brilliant ideas in between.
Second the overall claim is made for accelerating science in general, but examples cited are primarily for soft or biological sciences. Where it might be difficult to reason from first principles, I can certainly see data analytics helping build phenomenological models which can have useful descriptive value, but less clearly predictive value which generally has to rest on a deeper understanding than available data can provide.
SO, tool to help the pursuit of science, absolutely. I am concerned though about where we draw the boundary between helping science with analytics and doing science with analytics.
Thanks. You too also usually bring very nice articles from in-depth sources, much better than what goes for journalism these days.
As for your contentions:
>> First the approach helps purely with empirical science.
One of the approaches mentioned in this study is "literature based discovery". In the simple form it's trying to find hidden connections between two different unrelated fields by noticing sometimes they talk about similar phenomena, via analyzing many articles and abstracts textually using machines.
Using this hidden connections as a starting point, could sometimes lead to a leap with further work, with some help from humans and machines. And often among humans, such hidden connections are excatly the creative phase you allude to.
As for your second claim, about first principles:
There's a software called nutonian.com that given a dataset , can find the SIMPLEST formula describing that dataset. So maybe when you're starting with the data set of mechanical data , you might get something like the equations of nweoton's 3 laws. And that's relatively close to first principles.
I've even seen an article in an ecology research journal, where they use this system to find math the describes the behavior of a specific ecosystem , and they believe this is something generically useful. If you like i can search for that article.
But sure, it's a new field, it's hard to tell where the machine limits are , and there might be big parts humans still need to do, altough personally i'm with the belief that we're not far from being eclipsed intellectually by machines.
