[mlpack] Fwd: GSoC Project: LMNN (via LRSDP) & BoostMetric Implementation

[Ryan Curtin]

On Fri, May 11, 2018 at 08:00:37PM +0530, Manish Kumar wrote:
> Hi Ryan,
> 
> Thanks for looking into this.
> 
> Here is the link <https://code.google.com/archive/p/boosting/downloads> to
> original matlab implementation, I have compared with.The above idea
> suggested by you is similar to what BoostMetric's authors had implemented.
> Comparing with it, I saw that single pass KNN is slightly slower than
> multipass on lower value of k (though I haven't check on large datasets,
> which could be a game changer. I will check this on a large dataset),
> whereas on large values of k (say k=20), single pass KNN outperforms
> multipass one significantly( even on 125 points multipass has a running
> time of over 5 secs).
> 
> One more thing, that I think is worth disscussing about, is the number of
> constraints. From Boostmetric's implementation we will always be getting k
> * k * N number of constraints, whereas by restricting the impostors
> constrained within the boundary of kth targetNeighbor, we can reduce
> constraints by some amount. Let me know, what you think.

Sorry for the slow response---I think that most of this discussion
happened in IRC this morning, but I do want to at least point out that I
wouldn't expect the original MATLAB implementation to be a very good
benchmark.  In many cases MATLAB code can be written in such a way that
it runs very slowly.

-- 
Ryan Curtin    | "I'm so likable!"
ryan at ratml.org |  - Frank