I watched this great talk given by Daniel Wolpert the other day, titled The real reason for brains. It included some stuff on Bayesian inference that are pretty relevant with what we're doing at the university, especially MLPR. I hope you like it:
Great idea to start sharing talks here! I watch quite a lot of TED videos, but hadn't seen this one before.
I'm very interested in neuroscience, and I can tell you this bayesian viewpoint is really catching on big time in the neurosciences. It really explains a lot of phenomena very elegantly.
If you are interested in this kind of stuff, I highly recommend the journal of neural computation: http://www.mitpressjournals.org/loi/neco It publishes research on the intersection of machine learning and neuroscience.
I believe this fusion of brain research and AI is very promising, both for understanding the brain and building intelligent computers. From the neuroscience perspective, we have to move beyond just "butterfly collecting": collecting tons and tons of facts on what gene gets expressed where and what neurotransmitter has what effect on membrane potentials and so on.. What's needed is to embed these facts in a strong theory that explains things on a more abstract level.
As for neuroscience informing AI research, we have to keep in mind the analogy that "planes are not birds"; to build intelligent machines, we don't necessarily have to mimic nature. Nevertheless, our brain is the only example of true intelligence that we have, so it makes sense to reverse engineer it. And guess what.. we CAN fly like a bird, and it's beautiful too: http://www.ted.com/talks/a_robot_that_flies_like_a_bird.html
I thought it would be nice to extend our horizons a bit beyond what the university offers. Thanks for the journal by the way. Can you suggest some seminal papers from it which I can read and see if it's for me?
About "butterfly collecting", I think that a lot of disciplines have a serious deficiency in "being scientific". Social sciences have just recently started to discover statistics, for instance. So, it might take a while, but the wonders of mathematics and the tools computer scientists build based on them will eventually reach and inform all aspects of human inquiry.
As for neuroscience informing computer science, I don't see mimicking as something that should or should not be avoided; instead I think it should always be the first step. By trying to mimic nature we learn something that is fundamental to the understanding of the underlying concept: Take perceptron for example: would machine learning be where it is today if people didn't think to simply mimic their (now outdated) understanding of neurons? I doubt it.
"You insist that there is something a machine cannot do. If you tell me precisely what it is a machine cannot do, then I can always make a machine which will do just that." John von Neumann
I'll tell you precisely what a machine cannot do: It cannot decide with certainty whether a given program finishes running or continues to run forever.
Now, make me a machine that can do just that, Mr. Neumann. :p
I can write such a program, but it might run forever ;)
Since we're quoting famous physicists, here's one by Feynman: "What I cannot create, I do not understand." (On his blackboard when he died) Very true of the brain as well.
Not sure if that was a serious question or not, but consider this:
A program that runs forever if a certain number theoretic proposition is true, and halts otherwise. There exist number theoretic propositions that are true (cannot find a counterexample), but that are not provable (see Godels incompleteness theorem). So human or computer has to answer an undecidable question (no proof for or against exists), so it cannot answer.
This has nothing to do with actual programming practice of course :)
Great idea to start sharing talks here! I watch quite a lot of TED videos, but hadn't seen this one before.
ReplyDeleteI'm very interested in neuroscience, and I can tell you this bayesian viewpoint is really catching on big time in the neurosciences. It really explains a lot of phenomena very elegantly.
If you are interested in this kind of stuff, I highly recommend the journal of neural computation: http://www.mitpressjournals.org/loi/neco
It publishes research on the intersection of machine learning and neuroscience.
I believe this fusion of brain research and AI is very promising, both for understanding the brain and building intelligent computers. From the neuroscience perspective, we have to move beyond just "butterfly collecting": collecting tons and tons of facts on what gene gets expressed where and what neurotransmitter has what effect on membrane potentials and so on.. What's needed is to embed these facts in a strong theory that explains things on a more abstract level.
As for neuroscience informing AI research, we have to keep in mind the analogy that "planes are not birds"; to build intelligent machines, we don't necessarily have to mimic nature. Nevertheless, our brain is the only example of true intelligence that we have, so it makes sense to reverse engineer it. And guess what.. we CAN fly like a bird, and it's beautiful too: http://www.ted.com/talks/a_robot_that_flies_like_a_bird.html
I thought it would be nice to extend our horizons a bit beyond what the university offers. Thanks for the journal by the way. Can you suggest some seminal papers from it which I can read and see if it's for me?
ReplyDeleteAbout "butterfly collecting", I think that a lot of disciplines have a serious deficiency in "being scientific". Social sciences have just recently started to discover statistics, for instance. So, it might take a while, but the wonders of mathematics and the tools computer scientists build based on them will eventually reach and inform all aspects of human inquiry.
As for neuroscience informing computer science, I don't see mimicking as something that should or should not be avoided; instead I think it should always be the first step. By trying to mimic nature we learn something that is fundamental to the understanding of the underlying concept: Take perceptron for example: would machine learning be where it is today if people didn't think to simply mimic their (now outdated) understanding of neurons? I doubt it.
"You insist that there is something a machine cannot do. If you tell me precisely what it is a machine cannot do, then I can always make a machine which will do just that." John von Neumann
ReplyDeleteI'll tell you precisely what a machine cannot do: It cannot decide with certainty whether a given program finishes running or continues to run forever.
ReplyDeleteNow, make me a machine that can do just that, Mr. Neumann. :p
I can write such a program, but it might run forever ;)
ReplyDeleteSince we're quoting famous physicists, here's one by Feynman: "What I cannot create, I do not understand." (On his blackboard when he died) Very true of the brain as well.
I'll see if I can dig up some papers later on.
Humans can decide whether a program finishes running or continues to run forever, so why can't a computer program do that? ^_^
ReplyDeleteNot sure if that was a serious question or not, but consider this:
ReplyDeleteA program that runs forever if a certain number theoretic proposition is true, and halts otherwise. There exist number theoretic propositions that are true (cannot find a counterexample), but that are not provable (see Godels incompleteness theorem). So human or computer has to answer an undecidable question (no proof for or against exists), so it cannot answer.
This has nothing to do with actual programming practice of course :)