[mlpack] Contribute to Manifold learning based on spectral methods
Stephen Tu
tu.stephenl at gmail.com
Thu May 7 14:18:56 EDT 2015
Great question!
A good sanity check is to actually generate data from a manifold (like a
swiss roll or a cicle) and make sure the algorithm can indeed "unroll" the
structure. But in general testing is tricky, b/c there are two sources of
error: (a) error that is fundamental to the technique, and (b) error of
your implementation. You could of course also write unit tests to ensure
each piece of your implementation is correct (e.g. you are indeed creating
the adjacency matrix correctly and you are indeed solving the eigenvalue
problem correctly). But other than this there's not too much you can really
do I suppose. If there was, I think the authors would have included such an
evaluation in the paper.
Perhaps Ryan might have some good suggestions.
On Thu, May 7, 2015 at 2:09 AM, Shangtong Zhang <zhangshangtong.cpp at qq.com>
wrote:
> Hi Stephen,
>
>
> I want to contribute to the idea Manifold learning based on spectral
> methods.
>
> I implement LLE, LE, MDS and ISOMAP within the framework referred in the
> paper for offline points.
>
> It seems many lib plot the embedding points to prove the algorithm is
> right.
>
> Do you have some idea about how to write tests to prove the correctness of
> the algorithm.
>
>
> Best
>
> Regards
>
> Shangtong Zhang,
> Third Year Undergraduate,
> School of Computer Science,
> Fudan University, China.
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://mailman.cc.gatech.edu/pipermail/mlpack/attachments/20150507/84aa2df3/attachment.html>
More information about the mlpack
mailing list