Marginal MLE/MPD estimates from Mamba/Lora (GSOC?) (Crosspost)

Previous Topic Next Topic
 
classic Classic list List threaded Threaded
3 messages Options
Reply | Threaded
Open this post in threaded view
|

Marginal MLE/MPD estimates from Mamba/Lora (GSOC?) (Crosspost)

Jameson Quinn
I'm interested in working on a native Julia statistical modeling package (ie, Mamba or Lora) to add other methods. That would start (and, frankly, probably end) with optimization (with some ideas I have for automatically quasi-marginalizing latent variables, so that optimization is actually useful in nontrivial models). It might eventually include other turnkey stuff (variational Bayes, particle methods?). If Stan, with its slow compile times and monolithic development paradigm, can be a flexible modeling tool that allows quick switching between fully-Bayesian models and quicker approximations, then Julia, with its inherent flexibility and prettier syntax, should be able to do the same.  

(The great thing about using Julia for MCMC is the flexibility. For instance: Want to enable turnkey simulation studies, where you build a model and press a button and it does all the data-generation and parameter-fitting to give you nice graphs of convergence diagnostics? If you're using Stan, good luck. If you're in Julia, this is a SMOP.)

So, several questions:

0. Am I correct that neither Mamba or Lora does anything but MCMC at the present, and that there is no easy way to take a model in the correct format for either of these packages and (say) get an MLE/posterior mode through some optimizing package?
1. Would this make an appropriate GSoC project? It isn't really general language/packaging features, but in my opinion there is a real use case for it; it could be the Julia "killer app" that would get people to switch over from R. (In one fell swoop, you'd have a tool that could potentially compete with dozens of different R packages, from hierarchical linear modeling to Gaussian processes and more).
2. Is this the right place to ask these questions?

--
You received this message because you are subscribed to the Google Groups "julia-stats" group.
To unsubscribe from this group and stop receiving emails from it, send an email to [hidden email].
For more options, visit https://groups.google.com/d/optout.
Reply | Threaded
Open this post in threaded view
|

Re: Marginal MLE/MPD estimates from Mamba/Lora (GSOC?) (Crosspost)

Benjamin Deonovic
Take a look at the Mamba tutorial and extensive documentation (http://mambajl.readthedocs.org/en/latest/) if you haven't already. 

0. Not sure what the question is. Both packages do MCMC. MCMC gives you samples from the posterior. Mamba does a few other things necessary for MCMC (plotting, convergence diagnostics, etc)
1. Sounds like a good project to me!
2. Yeah!

On Wednesday, March 2, 2016 at 2:54:46 PM UTC-6, Jameson Quinn wrote:
I'm interested in working on a native Julia statistical modeling package (ie, Mamba or Lora) to add other methods. That would start (and, frankly, probably end) with optimization (with some ideas I have for automatically quasi-marginalizing latent variables, so that optimization is actually useful in nontrivial models). It might eventually include other turnkey stuff (variational Bayes, particle methods?). If Stan, with its slow compile times and monolithic development paradigm, can be a flexible modeling tool that allows quick switching between fully-Bayesian models and quicker approximations, then Julia, with its inherent flexibility and prettier syntax, should be able to do the same.  

(The great thing about using Julia for MCMC is the flexibility. For instance: Want to enable turnkey simulation studies, where you build a model and press a button and it does all the data-generation and parameter-fitting to give you nice graphs of convergence diagnostics? If you're using Stan, good luck. If you're in Julia, this is a SMOP.)

So, several questions:

0. Am I correct that neither Mamba or Lora does anything but MCMC at the present, and that there is no easy way to take a model in the correct format for either of these packages and (say) get an MLE/posterior mode through some optimizing package?
1. Would this make an appropriate GSoC project? It isn't really general language/packaging features, but in my opinion there is a real use case for it; it could be the Julia "killer app" that would get people to switch over from R. (In one fell swoop, you'd have a tool that could potentially compete with dozens of different R packages, from hierarchical linear modeling to Gaussian processes and more).
2. Is this the right place to ask these questions?

--
You received this message because you are subscribed to the Google Groups "julia-stats" group.
To unsubscribe from this group and stop receiving emails from it, send an email to [hidden email].
For more options, visit https://groups.google.com/d/optout.
Reply | Threaded
Open this post in threaded view
|

Re: Marginal MLE/MPD estimates from Mamba/Lora (GSOC?) (Crosspost)

Jameson Quinn

0. Not sure what the question is. Both packages do MCMC. MCMC gives you samples from the posterior. Mamba does a few other things necessary for MCMC (plotting, convergence diagnostics, etc)

For example: Stan does MCMC, variational Bayes, or posterior likelihood optimization (BFGS). I was asking if Mamba did the latter. It would appear it doesn't. That's OK; that would be part of my proposal then.

--
You received this message because you are subscribed to the Google Groups "julia-stats" group.
To unsubscribe from this group and stop receiving emails from it, send an email to [hidden email].
For more options, visit https://groups.google.com/d/optout.