Hello Yaowen, > Currently I am doing research in the area of Deep Reinforcement Learning. > When I > am looking for a cool project for GSoC 2017, the "Reinforcement Learning" > project on the mlpack github wiki really caught my attention.
Great that you like the idea, let me know if you need any starting points. > Actually, I read lots of recent papers about reinforcement learning, and work > on > models like DQN, A3C with my partners in the lab using tensorflow. So I think > this project is really cool and just meets what I am working on. I wonder that > if mlpack can get a better performance compared to tensorflow and other > frameworks on a multi-core CPU. We are working on it :) > Now I am diving into the mlpack codebase and have a question. Recently, lots > of > works in papers have complex models which consist multiple inputs and outputs. > And they may have multiple optimization stages when training them. In each > stage, the algorithm trains a different set of parameters and some parameters > are regarded as constants. In the meanwhile, some layers in the models could > be > recurrent neural networks, which need to reset the state from time to time. > Since I am not familiar with the codebase and all API related to ann, I am > curious about how to handle all of these? > > Currently I only found that one can use Forward and Backward to handle complex > training schemes manually. But as for high-level APIs, I guess FNN and RNN are > not capable to handle these things? So, that's basically a question about how you design the model, and I think the FFN and RNN classes can handle some situations but definitely not all, e.g. lets say you have a sequential network and a recurrent network and both get different or the same input you can simply concat both models using the FFN class. The training of the recurrent network is completely separated from the sequential model you can even share the weights between both models. But as I said there are situations where you have to implement another strategy where the existing classes can't be used. > BTW, as already mentioned by others, adding the ELU activation function is > really a good idea. :) Any contributions of new techniques or efficiency improvements for existing implementations are always welcome, so if you have some ideas let us know. I hope this is helpful, let us know if you have any more questions. Thanks, Marcus > On 28 Feb 2017, at 14:47, Yao-Wen Mao <[email protected]> wrote: > > Hello everyone, > > I am Yaowen, a fourth year undergraduate student studying Electrical > Engineering at National Taiwan University. Programming is my interest. I like > to investigate frameworks and features of programming languages, and create > some fun projects with them. > > Currently I am doing research in the area of Deep Reinforcement Learning. > When I am looking for a cool project for GSoC 2017, the "Reinforcement > Learning" project on the mlpack github wiki really caught my attention. > > Actually, I read lots of recent papers about reinforcement learning, and work > on models like DQN, A3C with my partners in the lab using tensorflow. So I > think this project is really cool and just meets what I am working on. I > wonder that if mlpack can get a better performance compared to tensorflow and > other frameworks on a multi-core CPU. > > Now I am diving into the mlpack codebase and have a question. > Recently, lots of works in papers have complex models which consist multiple > inputs and outputs. And they may have multiple optimization stages when > training them. In each stage, the algorithm trains a different set of > parameters and some parameters are regarded as constants. In the meanwhile, > some layers in the models could be recurrent neural networks, which need to > reset the state from time to time. > Since I am not familiar with the codebase and all API related to ann, I am > curious about how to handle all of these? > Currently I only found that one can use Forward and Backward to handle > complex training schemes manually. But as for high-level APIs, I guess FNN > and RNN are not capable to handle these things? > > BTW, as already mentioned by others, adding the ELU activation function is > really a good idea. :) > > Thanks, > Yaowen > _______________________________________________ > mlpack mailing list > [email protected] > http://knife.lugatgt.org/cgi-bin/mailman/listinfo/mlpack
_______________________________________________ mlpack mailing list [email protected] http://knife.lugatgt.org/cgi-bin/mailman/listinfo/mlpack
