Thanks! Could also an item to study and constrast the convergence of classical MLE / EM GMM with Bayesian GMM? I would like to check that they effectively convergence to the same solution when the number of samples grow. It would be interesting also to study their respective behaviors when facing singularities (e.g. small number of samples and comparatively large number of components). Bayesian GMM should be robust to singularities while this is not the case for classical MLE with EM.
We need to make sure to be user friendly in those cases: either by giving informative error message with potential strategies to solve the issues or automatically addressing if possible (e.g. random re-init of singular components). Finally I would like to have an example that shows how GMM can overfit by comparing likelihood on training and validation sets (normalized by the number of samples). We could for instance extend the BIC score example with a cross-validated likelihood plot as done in my notebook. -- Olivier 2015-06-16 20:18 GMT+02:00 Wei Xue <xuewe...@gmail.com>: > Hi all, > > I have updated the third blog post, in which I listed several steps I am > going to take in the next few weeks. > > Regards, > Wei Xue > > On Tue, Jun 9, 2015 at 6:35 PM, Wei Xue <xuewe...@gmail.com> wrote: >> >> Hi Olivier, Andreas, Loïc and group, >> >> I have finished the derivations of VBGMM and DPGMM with four types of >> covariance matrix. The draft includes the updating functions, lower bound >> and predictive distributions. >> >> Next, I will shift my focus to coding. >> >> Regards, >> Wei Xue >> >> On Tue, Jun 9, 2015 at 6:25 PM, Wei Xue <xuewe...@gmail.com> wrote: >>> >>> Anyway, thanks for your advice! I am OK with LaTeX now when using >>> snippets and auto-completion within my IDE. >>> >>> Wei >>> >>> On Sat, Jun 6, 2015 at 11:46 AM, Sturla Molden <sturla.mol...@gmail.com> >>> wrote: >>>> >>>> On 05/06/15 22:12, Wei Xue wrote: >>>> >>>> > I published a new post >>>> > >>>> > <http://xuewei4d.github.io/2015/06/05/gsoc-week2-vbgmm-and-gmm-api.html> >>>> > in my blog for the second week. Basically, I have done the derivation >>>> > of >>>> > VBGMM and working on cleaning GMM API. >>>> >>>> If you think typing LaTeX math is tedious, get MathType or MathMagic (or >>>> something similar). Even if you hand type parts of the equation, you can >>>> still copy&paste common symbols and expressions to make the typing >>>> faster. >>>> >>>> >>>> Sturla >>>> >>>> >>>> >>>> >>>> ------------------------------------------------------------------------------ >>>> _______________________________________________ >>>> Scikit-learn-general mailing list >>>> Scikit-learn-general@lists.sourceforge.net >>>> https://lists.sourceforge.net/lists/listinfo/scikit-learn-general >>> >>> >> > -- Olivier http://twitter.com/ogrisel - http://github.com/ogrisel ------------------------------------------------------------------------------ _______________________________________________ Scikit-learn-general mailing list Scikit-learn-general@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/scikit-learn-general
