Often pop-sci articles will talk about machine learning, particularly in the context of Big Data, as well as genetic algorithms, which are algorithms generated by creating a set of algorithms, testing them for solving some problem, and cross-breeding the most successful algorithms to create new, hopefully better algorithms. (I don’t understand how to structure algorithms so they can be crossed, but that’s neither here nor there.) But something not mentioned – perhaps because it’s too obscure – is the reward system. Not survival as a reward, but simply the reward for getting something right: resources.
So when the D-brief blog on Discover Magazine pointed me at Pigeons working as pathologists, my first reaction was how cool is this?
To train pigeon pathologists, a team led by Richard Levenson from the University of California Davis Medical Center and Edward Wasserman from the University of Iowa placed pigeons in a conditioning chamber fitted with a touch-screen monitor. Eight pigeons were shown 144 images of benign or malignant tissue samples at three levels of magnification, at different orientations and at different brightness levels. The birds made a diagnosis by pecking on a blue or yellow rectangle on the screen, and they received a tasty treat for each tissue sample they correctly identified.
And then I wondered. Do we reward algorithms when they get something right? And take the anti-thesis – suppose we have some N algorithms competing to get a “right answer”, and some M, less than N, get it right and are so rewarded – are the losing algorithms given the knowledge that they failed? Are these algorithms – and we might as well call them entities at this point – aware of this failure and begin to employ more scarce resources to secure proper resources? And do the winners do the computer thing and immediately deploy the resources, or do they conserve them for later use? How do you conserve computer resources?
Just how much do we emulate Nature? And, to the extent that we don’t emulate Nature, what are we losing in our computerized emulations of Nature? Is it important to understand that an organism which succeeds at some test is then not pressured to develop a new way to solve the puzzle, unlike the loser that survives long enough for try #2?
I think I’ve done little more here than expose my ignorance of artificial intelligence and/or biological simulations.
