On Tue, Mar 28, 2017 at 03:41:34AM +0530, Sidakpal Singh wrote: > Hi, > > I am a final year CS undergraduate at Indian Institute of Technology (*IIT*) > Roorkee and will be joining *Microsoft Research* later this fall. I am > interested in being a part of GSoC with mlpack and contribute to the* > Essential Deep Learning Modules* project. > > I believe that Generative Adversarial Networks are an extremely hot topic > of research in Deep learning these days. So, it would be immensely useful > if we are able to provide an API of the major types of GANs. Here's what I > have in mind. > > 1. The *vanilla GAN* > 2. Generative Moment Matching Networks (*GMMN*) > 3. Deep Convolutional Generative Adversarial Nets (*DCGAN*) > 4. *Conditional GANs* (Basic, StackGANs, Plug and Play Generative Networks) > 5. *InfoGAN* (encode meaningful features in the noise Z) > 6. *Wasserstein *GAN & f-GAN (*Divergence minimization*) > > The above are the most commonly used variants of GANs and the ones against > which > researchers *generally compare* their variant of GANs. Since this will > broadly cover the major > GAN types, it would allow for *easy extensibility* when somebody wants to > design their own variant. Further, as RBMs take much *more time for > sampling*, providing these variants might be sufficient. Lastly, I am open to > *discussing other models* we may include which I might have missed and are > important. > > Btw, if time permits, we may also include *Variational & Adversarial > Autoencoders* or *PixelRNN* and some applications like image to image > translation, inpainting etc. > > *Background: *Since last summer I have been working on *formulating a new > loss function for training GANs* based on Optimal Transport distances. This > work started off via my internship at *Kyoto University*, Japan with *Prof. > Marco Cuturi*. This has given me a great experience in dealing with the > instability issues involved in training GANs. Further, I have a very solid > understanding of *Wasserstein distances* which are essentially a kind of > Optimal transport distances. Also, here is a link to an implementation of > GAN in Chainer, which I built for playing around with them. > > https://github.com/sidak/GAN-Chainer > > I have a strong background in ML through courses and projects, and have > also used TensorFlow and Shogun. Besides, I have also carried out research > at *Xerox* and *Purdue University*, USA and my work has also been published > in *IJCAI*. > > It would be really great if you can *share your views* on this and suggest > if this would *serve as a * > *good plan* for the GSoC. Also, it will be really wonderful if you can > guide me on to the next steps. :)
Hi Sidak, Thanks for getting in touch. I think that implementing GANs and related techniques could be an interesting and exciting summer project. If you're planning to prepare a proposal on that, I would suggest that you spend a good amount of time with the mlpack ANN codebase so that you can detail, in your proposal, how the different GAN modules will be implemented. It's really important for us, when we evaluate proposals, to be able to see how well a proposed piece of code will fit into the rest of the library. I hope that this is helpful; let me know if I can clarify anything. Thanks, Ryan -- Ryan Curtin | "Get out of the way!" [email protected] | - Train Engineer Assistant _______________________________________________ mlpack mailing list [email protected] http://knife.lugatgt.org/cgi-bin/mailman/listinfo/mlpack
