Regarding Deep Learning Algorithms in GraphLab and GraphChi

User 32 | 3/18/2014, 6:49:42 PM

First of all thank you very much for the amazing work done by GraphLab team for all of machine learning community and making GraphLab and GraphChi open-source.

I wanted to inquire about the implementation of Deep Learning algorithms in Graphlab. In the future supported features webpage of Graphlab create "" it is mentioned that Deep Learning implementation is coming soon. I was curious to know which all deep learning algorithms do you have plans to make available in future releases of GraphLab create.

I have been seen the implementation of Restricted Boltzmann Machine (RBM) code in collaborative filtering module of GraphChi. Is there also implementation of other Deep Learning Methods in GraphChi or GraphLab.

Actually, I have some experience working in Deep Learning in Computer vision and Speech Processing mainly in Matlab and Python. I have to start Deep Learning work for Overlapping Community Detection using GraphChi and GraphLab open source codes. Could you please give me some guidelines and advice to be followed for extension and implementing Deep Learning codes in GraphLab/GraphChi.

Thank you so much for your time and effort.

Best Regards, Devendra


User 6 | 3/19/2014, 5:59:14 AM

Hi Devendra, As you mentioned correctly we do have deep learning on our roadmap, and it is true we have RBM implemented as part of GraphChi. We are looking into adding additional functionality to GraphLab Python but we can not commit to a date yet. It will help us if you and our users could give us some recommendations about the most useful methods to add and we will consider accordingly.

User 32 | 3/19/2014, 6:23:39 AM

Hi Danny,

Thanks so much for the help. Another thing I wanted to take advice on is regarding implementing deep learning methods on top of GraphLab/GraphChi. What all are the the design things which we need to follow in order to gain maximum speed boost up in Graphlab/Graphchi while implementing deep learning methods.

User 6 | 3/19/2014, 6:32:58 AM

Hi Devendra, You question is too general to answer clearly... Basically GraphLab open source fits very well algorithms that are: 1) iterative - algorithm has to make a few passes on the data 2) there are sparse data/parameter dependencies, namely the problem can be represented as a graph 3) each update is concerned with a local neighborhood of the graph You will need to examine the algorithm you want to implement and see how well it fits into the above.

User 32 | 3/20/2014, 9:53:25 AM

Thanks so much Danny for the useful tips. I will ask more once I start implementing and experimenting different modules.