Re: your former postdocs remark about whether Newton could've discovered his theory of gravity via machine learning, the answer's no:
http://errorstatistics.com/2013/03/04/big-data-or-pig-data/ (click through, you'll get a kick out of it)
I was a little surprised that your piece didn't include some nods towards some of the theoretical evolutionary biologists working on these issues. I'm thinking of folks like Michael Lynch, Gunter Wagner, and Steven Frank. Do molecular biology/genomics folks ever look at that stuff? If not, why not? If so, what do they think of it? Honest questions, definitely not trying to imply that the molecular types are all ignorant or that evolutionary biologists have already figured all this stuff out or anything like that.
Haha, the pig-data post is hilarious! A very succinct and incisive point. I think I used to think that way, and still do for the most part, but let me put forth the following. If we accept that our minds do not derive from the divine and are in fact nothing more than highly complex, differently designed computers, then the deductive reasoning of Newton et al. is indeed just machine learning, albeit of a kind we have not yet remotely achieved in silicon. So formally, it *is* possible for machine learning to discover gravity. I think this is not just an academic point but an increasingly practical one. If we accept that it is in principle possible for a computer to make the sort of deductions we now consider science, then it is also possible, perhaps probable, that those computers may quickly thereafter be able to understand systems of complexity that our "low bandwidth" communication systems are unable to cope with. I think this may require a rethinking of what we mean by "understand" and "learn". Is it really the logic that we aim to communicate, or is it the intuition that we seek to transfer? If the latter, then perhaps our standard definition of what it means to understand or derive Newtonian mechanics may need a rethinking.
Ach, and yes, good point about the evolutionary biologists. We totally meant to cite Wagner and just plain forgot, ugh–less familiar with the other two, but I don't get out very much (though Lynch has been on my desk for some time awaiting a reading). I think that mol. bio and genomics types do know about this line of thought, perhaps some more than others. I think I saw a few papers in which Wagner essentially argued that modularity is essentially a given at this point. I'm not sure I agree with that, but I definitely agree with a lot of the ideas about pleiotropy. I do think that molecular biologists would do well to take some time to consider these questions more deeply in their own work.
Meanwhile, back to writing a super boring grant... :)
Very nice article. In regards to Newton's Laws, I think it is possible to automate the discovery of analytical laws (see Schmidt and Lipson, Science, 2009).
Nice piece, seems quite cogent to me.
ReplyDeleteRe: your former postdocs remark about whether Newton could've discovered his theory of gravity via machine learning, the answer's no:
http://errorstatistics.com/2013/03/04/big-data-or-pig-data/ (click through, you'll get a kick out of it)
I was a little surprised that your piece didn't include some nods towards some of the theoretical evolutionary biologists working on these issues. I'm thinking of folks like Michael Lynch, Gunter Wagner, and Steven Frank. Do molecular biology/genomics folks ever look at that stuff? If not, why not? If so, what do they think of it? Honest questions, definitely not trying to imply that the molecular types are all ignorant or that evolutionary biologists have already figured all this stuff out or anything like that.
Haha, the pig-data post is hilarious! A very succinct and incisive point. I think I used to think that way, and still do for the most part, but let me put forth the following. If we accept that our minds do not derive from the divine and are in fact nothing more than highly complex, differently designed computers, then the deductive reasoning of Newton et al. is indeed just machine learning, albeit of a kind we have not yet remotely achieved in silicon. So formally, it *is* possible for machine learning to discover gravity. I think this is not just an academic point but an increasingly practical one. If we accept that it is in principle possible for a computer to make the sort of deductions we now consider science, then it is also possible, perhaps probable, that those computers may quickly thereafter be able to understand systems of complexity that our "low bandwidth" communication systems are unable to cope with. I think this may require a rethinking of what we mean by "understand" and "learn". Is it really the logic that we aim to communicate, or is it the intuition that we seek to transfer? If the latter, then perhaps our standard definition of what it means to understand or derive Newtonian mechanics may need a rethinking.
DeleteAch, and yes, good point about the evolutionary biologists. We totally meant to cite Wagner and just plain forgot, ugh–less familiar with the other two, but I don't get out very much (though Lynch has been on my desk for some time awaiting a reading). I think that mol. bio and genomics types do know about this line of thought, perhaps some more than others. I think I saw a few papers in which Wagner essentially argued that modularity is essentially a given at this point. I'm not sure I agree with that, but I definitely agree with a lot of the ideas about pleiotropy. I do think that molecular biologists would do well to take some time to consider these questions more deeply in their own work.
Meanwhile, back to writing a super boring grant... :)
Very nice article. In regards to Newton's Laws, I think it is possible to automate the discovery of analytical laws (see Schmidt and Lipson, Science, 2009).
ReplyDeleteThanks! Indeed, we make exactly this point and cite this paper later in the article.
Delete