Ok, today I will start to read SGD code which is in the repo and I will
think about how to implement AGSD nicely. Your tricks are really useful.
On 17 Nov 2011 18:16, "Ted Dunning" wrote:
The key tricks are:
- do the updates of the averaged model in a sparse fashion. This will
require doubling the space kept by the model
- determine when to switch to averaging
In addition we should bring in at the same time
- more flexibility on loss function (to allow the code to implement SVM
Hi Ted;
I start to read the paper and I think I will finish it today. It is a quite
nice approach and
thanks for supervision.
Cheers
Ürün
On Wed, Nov 16, 2011 at 8:14 PM, Ted Dunning wrote:
> On Wed, Nov 16, 2011 at 9:50 AM, urun dogan wrote:
>
> >
> > I have written the previous email before
On Wed, Nov 16, 2011 at 9:50 AM, urun dogan wrote:
>
> I have written the previous email before reading Josh's email. Are there
> any objections if I conclude that: implementation of SGD/ASGD based methods
> have priority and therefore I will start implement these methods soon ?
>
I think that t
Regarding linear classifiers, I think that the cluster/classifier
unification and introduction of ASGD are the only items of substantial
impact.
On Wed, Nov 16, 2011 at 9:39 AM, Josh Patterson wrote:
> Could you then make a list of JIRAs that you think are more
> interesting in the near term, po
kernels, parallelize the algorithm and we can
>> have
>> > a online SVM method for large/web scale datasets.
>> >
>>
>> Now this begins to sound right.
>>
>> Honestly I am so much into SVM and kernel machines and I fear that I am
>> > making bi
am
> > making big fuss out of small problems.
>
>
> My key question is whether you have problems that need solving. Or do you
> have an itch to do an implementation for the sake of having the
> implementation?
>
> Either one is a reasonable motive, but the first is prefer
I'd have to admit my interest in SVMs is more of the "abstract
curiosity" nature;
In the case of needed focus in the near term, similar to how Grant tagged:
https://issues.apache.org/jira/secure/IssueNavigator.jspa?reset=true&jqlQuery=labels+%3D+MAHOUT_INTRO_CONTRIBUTE
Could you then make a list
On Wed, Nov 16, 2011 at 12:09 AM, urun dogan wrote:
> Hi All;
>
> As I mentioned, I really found interesting to implement SGD and Pegasos. We
> can add Pegasos into SGD modules.
Based on Leon Bottou's results, I would recommend a simple SGD
implementation of SVM rather than Pegasos.
http://leo
Hi All;
As I mentioned, I really found interesting to implement SGD and Pegasos. We
can add Pegasos into SGD modules. However, I think there are two issues we
need to clarify:
1) In general SGD like ideas are used for online learning (of course they
can be converted to batch learning) and Pegasos
Hi Urun and Josh,
I'd also be interested in helping out in whatever way I can.
One question, I've noticed that MAHOUT-334 was not ultimately adopted. Do we
know the reason for this?
Would it be best to finish out the patch in 232, or instead add the
functionality into the existing SGD modules
Urun,
I've been looking at MAHOUT-232 and reading Nello Cristianini's book
on SVMs. It sounds like you've done considerable more work than I in
this arena. I'd be interested in collaborating with you on finishing
out this patch, if you are interested in that type arrangement (there
is plenty of wor
Dear Josh and Ted;
Both ideas are very attractive. Honestly I want to do both of them. I am
completely aware
that this quite some work to do. As I mentioned before, I am a Postdoc now
and I am trying
to develop new techniques by using AGSD. During my PhD I developed an
efficient solver for
multicl
ASGD is also an opportunity laying on the table.
http://leon.bottou.org/projects/sgd
It would be lovely to have the current SGD system upgraded to use ASGD and
allow multiple loss functions to allow SVM training as well as the current
logistic regression. I would be happy to supervise, but can't
Urun,
Sounds like you have quite a bit of SVM experience. There is always:
https://issues.apache.org/jira/browse/MAHOUT-232
to take a look at which involves getting SVMs going in Mahout. I've
looked at it a bit while working on some smaller patches, I'd be
interested in discussing it with you giv
https://cwiki.apache.org/confluence/display/MAHOUT/How+To+Contribute has some
tips, ideas, etc. It's usually best to start with a few small patches to get
your feet wet w/ the development process.
On Nov 14, 2011, at 7:44 PM, Raphael Cendrillon wrote:
> Hi Urun,
>
> I'm in a very similar s
Hi Urun,
I'm in a very similar situation. I have a background (PhD) in optimization and
signal processing and some experience with principal component analysis. I'm
fairly comfortable with Java.
I'm also very interested in Mahout, and large scale problems.
If we can find a suitable area I wo
Hi All;
I want to give my congratulation to all of the contributors of the project.
I found the idea of this project so nice and I want to contribute to the
project.
I am postdoctoral researcher who is involved on developing machine learning
algorithms. During my PhD I have developed several mult
18 matches
Mail list logo