I’m quite new to LW, and find myself wondering whether Hidden Markov models (HMM) are underappreciated as a formal reasoning tool in the rationalist community, especially compared to Bayesian networks?
Perhaps it’s because HMM seem to be more difficult to grasp?
Or it’s because formally HMM are just a special case of Bayesian networks (i.e. dynamic Bayes nets)? Still, HMM are widely used in science on their own.
For comparison, Google search “bayes OR bayesian network OR net” site:lesswrong.com gives 1,090 results.
Google search hidden markov model site:lesswrong.com gives 91 results.
Well, Kurzweil is an extremely accomplished inventor aside from being a pie-in-the-sky futurist, so when he says something about a particular algorithm working well, I assume he knows what he’s talking about. He seems to think hidden hierarchical Markov models are the best way to represent the hierarchical nature of abstract thought.
I’m not saying he’s correct, just saying, it seems to be a popular idea.
There’s a proliferation of terminology in this area; I think a lot of these are in some sense equivalent and/or special cases of each other. I guess “Bayesian network” is more consistent with the other Bayes-based vocabulary around here.
I’m quite new to LW, and find myself wondering whether Hidden Markov models (HMM) are underappreciated as a formal reasoning tool in the rationalist community, especially compared to Bayesian networks?
Perhaps it’s because HMM seem to be more difficult to grasp?
Or it’s because formally HMM are just a special case of Bayesian networks (i.e. dynamic Bayes nets)? Still, HMM are widely used in science on their own.
For comparison, Google search “bayes OR bayesian network OR net” site:lesswrong.com gives 1,090 results.
Google search hidden markov model site:lesswrong.com gives 91 results.
Hidden Markov models are a reasoning model to solve a specific problem. If you don’t face that specific problem they are no use.
Most of the problems we discuss aren’t modeled well with HMMs.
Out of curiosity, did you happen to read Kurzweil’s recent book on HHMMs?
I think the safest answer is that a HMM is just a specific way of mathematically writing down an updating Bayesian network.
No, never heard of it. I’m not an Utopian, and from what I know about Kurzweil’s ideas and arguments, they don’t seem to be sound enough.
Well, Kurzweil is an extremely accomplished inventor aside from being a pie-in-the-sky futurist, so when he says something about a particular algorithm working well, I assume he knows what he’s talking about. He seems to think hidden hierarchical Markov models are the best way to represent the hierarchical nature of abstract thought.
I’m not saying he’s correct, just saying, it seems to be a popular idea.
There’s a proliferation of terminology in this area; I think a lot of these are in some sense equivalent and/or special cases of each other. I guess “Bayesian network” is more consistent with the other Bayes-based vocabulary around here.