Hello Michael, if you are willing to implement A3C or DQN if there is time left or along the way as you proposed I think this could be a really interesting project. Let us know what you think.
Thanks, Marcus > On 24 Mar 2017, at 22:26, Michael H Gump <[email protected]> wrote: > > Hi Marcus, > > My thought process was that aspects of the PathNet architecture are > significant and independent of RL. In their paper they apply it to > classification tasks as well such as MNIST and I was envisioning a module > that would easily allow for training a single network on multiple tasks in > the same manner as PathNet. The module would stand on its own and be designed > to work the with the rest of the ANN portion of mlpack. Would a reasonable > proposal be to design this module and test it with various parts of mlpack’s > existing codebase (i.e. on MNIST, other classification tasks)? Possibly > implementing A3C or DQN along the way as another avenue to use the module? > The deliverables would be the module, examples of how to use the module, and > examples of successful transfer learning between tasks. > > Best, > > Michael > > > >> On Mar 24, 2017, at 10:17 AM, Marcus Edel <[email protected] >> <mailto:[email protected]>> wrote: >> >> Hello Michael, >> >> thanks for your interest in the project. I like the idea, but as you already >> pointed out probably the most interesting part is to train PathNet in >> conjunction with A3C. Since there is no A3C implementation yet I'm not sure >> it's >> a good idea to create such a dependency. However, I think a somewhat >> reasonable >> idea would be to combine the implementation of the PathNet paper and the A3C >> method. Let us know what you think. >> >> Thanks, >> Marcus >> >>> On 23 Mar 2017, at 17:35, Michael H Gump <[email protected] >>> <mailto:[email protected]>> wrote: >>> >>> Hi MLPack, >>> >>> I’m new to the Github, the mailing list, and the mlpack project but I’ve >>> been going through the source code and the tutorials because I am very >>> interested in contributing to mlpack for GSoC. I had an idea for a GSoC >>> proposal but it’s a bit different from anything on the idea list so I >>> wanted to ask for feedback first. >>> >>> Recently, DeepMind released a paper called PathNet >>> (https://arxiv.org/pdf/1701.08734.pdf >>> <https://arxiv.org/pdf/1701.08734.pdf>) where they investigate fixing >>> evolution channels as a method for transferring learning between groups of >>> diverse tasks (Atari games). I think an interesting project could be to >>> develop the path fixing algorithms that allow PathNet to transfer learning. >>> I saw that Bang Liu had worked on NEAT in GSoC 2016 but I couldn’t find his >>> project so I’m not sure how much structure there is for neural evolution. I >>> was looking for feedback on how feasible this project could be in terms of >>> support and whether it was something that would be useful to mlpack. >>> >>> I also saw in the mailing list archives that a few people are already >>> interested in implementing DQN, A3C, etc. for GSoC 2017 and I think it >>> could be possible for me to collaborate with them (PathNet was run over A3C >>> in DeepMind’s tests so that is an obvious use case). But I think the >>> majority of this project would be independent of their work as it could >>> hopefully be designed to work with an arbitrary RL training technique. >>> >>> Also, I’m sorry if this was the wrong place to email this. I couldn’t find >>> a way to contact those working on the ANN modules directly. Let me know if >>> there’s a better place for me to ask for feedback on my idea. >>> >>> Best, >>> >>> Michael Gump >>> MIT Class of 2019 >>> _______________________________________________ >>> mlpack mailing list >>> [email protected] <mailto:[email protected]> >>> http://knife.lugatgt.org/cgi-bin/mailman/listinfo/mlpack >>> <http://knife.lugatgt.org/cgi-bin/mailman/listinfo/mlpack> >
_______________________________________________ mlpack mailing list [email protected] http://knife.lugatgt.org/cgi-bin/mailman/listinfo/mlpack
