Clatern GSoC 2015 Proposal

[Rinu Boney]

Hi all,

     I've already made some posts and discussions about this here. So, GSoC 2015 is here and I'm interested in applying. My idea is this - a Machine Learning library for Clojure. I've done some work on the project and the proposal. The code is available here - https://github.com/rinuboney/clatern and my proposal is available here - https://docs.google.com/document/d/1E3CrYw2HD1m6UgrqTdwlmoAw62kT97IdgK07s2M4d_c/edit?usp=sharing.

    Any criticism, feedback or suggestions regarding the project or the proposal is welcome.

Rinu.

 

[Mike Anderson]

I think working in this area would be a great GSoC project!

I've already provided you some thoughts via email. Would just like to re-emphasize that in order to make this a success the library needs to work "end to end" with other tools in the ecosystem. All the way from acquiring data sets into memory through to displaying results in Incanter etc. Might be a good idea to write an article / blog post or two showing how the complete process works.

[Rinu Boney]

I've been busy working on the proposal and I was thinking about writing a few blog posts. So, I've written one about multiclass classification. It shows usage of logistic regression and knn to classify the iris dataset. The post is in my blog here: https://rinuboney.github.io/2015/02/22/Multiclass-Classification-using-Clatern.html

[Mike Anderson]

Cool - link seems to be broken though?

[Rinu Boney]

Oops.. Here is the correct link: https://rinuboney.github.io/2015/02/26/Multiclass-Classification-using-Clatern.html

--
You received this message because you are subscribed to a topic in the Google Groups "Numerical Clojure" group.
To unsubscribe from this topic, visit https://groups.google.com/d/topic/numerical-clojure/S29eeB6qkcM/unsubscribe.
To unsubscribe from this group and all its topics, send an email to numerical-cloj...@googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

[Daniel Slutsky]

Thanks for this nice post!

For computing accuracy and other evaluation measures, you may like
this Java library:
https://github.com/kboyd/Roc
It is well documented (though it seems not so much in use yet).
The authours are very thoughtful and responsive.

> You received this message because you are subscribed to the Google Groups
> "Numerical Clojure" group.
> To unsubscribe from this group and stop receiving emails from it, send an

[Rinu Boney]

Thanks :-) The Roc library does seem well documented and underused. I'll definitely look into it!

[Rinu Boney]

Here is another blog post. It's a short tutorial on using KNN algorithm to classify handwritten digits. It's here: http://rinuboney.github.io/2015/03/01/digit-classification-knn.html

[Alex Ott]

you may also look to the Infer library (from Prismatic): https://github.com/alexott/infer - this is my fork updated to work with fresh clojure

With best wishes,                    Alex Ott
http://alexott.net/
Twitter: alexott_en (English), alexott (Russian)
Skype: alex.ott

[Rinu Boney]

Infer library looks good. It has lots of statistical and machine learning techniques. I could definitely refer to it for development of Clatern. I think it would have been better if it was developed to be well integrated with core.matrix and Incanter. I intend to use core.matrix and Incanter for linear algebra and statistical operations and focus on the machine learning part(models, learning algorithms, validation etc) in Clatern.

[Alex Ott]

Yes, I have the idea of merging parts of infer into Incanter...

[Matt Revelle]

Merging infer with Incanter was attempted about 4+ years ago and the respective authors decided against the approach. Updating infer to use core.matrix makes sense, but it's better to keep an ML library independent from Incanter. Prefer small libraries over tightly-coupled frameworks.

[Rinu Boney]

I completely agree Matt. I intend to build Clatern on top of core.matrix. By "well integrated with core.matrix and Incanter", I meant not to add Incanter as a dependency but to make it easy to use Incanter and Clatern together and to not reimplement the stuff that is already in Incanter. I meant something more like "well integrated with the ecosystem". Incanter is required to plot ROC curve, learning curves and other visualizations. Since both Clatern and Incanter 2.0 uses core.matrix, they should fit in well. It would be cool if Infer was ported to core.matrix.