[mlpack] Towards a robust deep learning module for FastML | Guidance for GSOC'16

[Marcus Edel]

Thanks for getting in touch and welcome!

> I have great interests in contributing to the FastML code base, particularly in
> extending or improving deep learning modules within FastML. In particular, I
> would like to implement GRUs and augment the current implementation of LSTMs to
> include peephole connections which have become very popular.

It would be great to see more deep learning modules, that said I would really
like to see a GRU implementation in the future. The LSTM layer already includes
an implementation of peephole connections.

> I would also like to implement the CBOW and the word2vec models for FastML as
> they have also become quite popular in the NLP community.

I think it doesn't make much sense to pick various methods an implement them,
without providing a model to use them. On the other side, I'm not sure that the
project is enough work for 3 months, in which you are expected to work full-time.
However, If you want to continue down that path, maybe we can
find something that fits better with the codebase.

I hope this is helpful, let me know if I can clarify anything.

Thanks, Marcus

> On 13 Mar 2016, at 07:04, Shikhar Murty <shikhar.murty at gmail.com> wrote:
> 
> Apologies, but I seem to have confused FastML for mlpack for some reason, due to the similar looking names. However, I meant mlpack.
> 
> 
> 
> On Sun, Mar 13, 2016 at 11:06 AM, Shikhar Murty <shikhar.murty at gmail.com <mailto:shikhar.murty at gmail.com>> wrote:
> Hi Marcus,
> 
> I'm a junior year EE undergrad (with elective concentration in CS) at IIT, Delhi. I'm an undergraduate researcher in Deep learning and NLP. I have implemented a bunch of ANNs and recently have started using FastML. I have familiarised myself with the code base of FastML, and understand the ANN implementation and the RNN implementation along with the optimisers module.  I'm quite keen on participating in GSOC'16 under FastML, and request you to mentor me for the same.
> 
> As an undergraduate researcher, I've read and understood multiple papers in Deep learning, for instance Collobert et al (NLP from scratch), Mikolov et al. (word2vec), Dropouts, Adam, Bengio's paper on neural network language models and so on.
> 
> I have great interests in contributing to the FastML code base, particularly in extending or improving deep learning modules within FastML. In particular, I would like to implement GRUs and augment the current implementation of LSTMs to include peephole connections which have become very popular. I would also like to implement the CBOW and the word2vec models for FastML as they have also become quite popular in the NLP community. 
> 
> Could you please give me some pointers on how I should structure my application for GSOC to boost my chances? 
> 
> Thank you for your time and I look forward to a positive response
> 
> Regards,
> 
> Shikhar Murty
> 

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://mailman.cc.gatech.edu/pipermail/mlpack/attachments/20160313/fe6cf484/attachment.html>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: smime.p7s
Type: application/pkcs7-signature
Size: 5233 bytes
Desc: not available
URL: <https://mailman.cc.gatech.edu/pipermail/mlpack/attachments/20160313/fe6cf484/attachment.bin>