Hi Rishabh,
> I wanted to update you regarding the project. Hopefully by tonight I will
> share with you the patch which incorporates refactoring in xapian-letor
> (insteadof xapian-core/letor). In the meanwhile I had discussed with
> Gaurav about the evaluation framework and he mentioned that Terrier(which
> is in JAVA) is considered good, so I went through its features and ead
> Andy's existing code for Xapian's evaluation which Olly shared
yesterday.
>
Sounds good. It makes sense that you and Gaurav share the same evaluation
framework so that at the end we do not end up with two different evaluation
modules.
Of course, in our case the input to the evaluation module would be
Letor::Ranklist. So, if you overload the method like evaluate() which
returns the NDCG at k or MAP supplying a Letor::Ranklist (in our case) and
Xapian::MSet (in Gaurav's case). You can coordinate among yourselves.
Basically we need to use it a lot. While training, we optimise a loss
function and validate the model using the same evaluation module.
> Also regarding Deep Learning I went through some a couple of papers on
> finding the structure of the deep network and will soon update the Notes
> section/blog with the details.
>
Great. Btw did you have look at the Matlab/Octave code written by Geoffrey
Hinton that is used in the Science paper. That's very informative and
understandable. Recently I played with it in some experiments and its quite
easy to follow.
Currently I am in Istanbul attending a conference so this week you might
not find me much around but most of the days I will try to answer e-mails
in the night.
Parth.
>
> Regards,
> Rishabh.
>
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.xapian.org/pipermail/xapian-devel/attachments/20120521/2cdcc425/attachment-0001.html>