Word embedding usage
angelo337 opened this issue · comments
hi there
is it possible to use DNC to train a classifier base on text?, encoding every work with an embedding of course.
thanks
@angelo337 good idea, it would be great to have the babi task as an example, similar to this implementation
The dnc core adheres to the standard TF rnn core interface so you can drop it into any training script where an rnn is trained on text. You can also wrap the core within another RNN core that contains embedding encoders & a softmax decoder. There are no plans to add a suite of tasks from our end but contributions are welcome!
The first few lines of this repo states: Any publication that discloses findings arising from using this source code must cite “Hybrid computing using a neural network with dynamic external memory"
I thought the point of this code was to publish the implementation outlined in the nature paper. If your intention was to publish this code and to get a peer and community review of your implementation and methodology as per the scientific method that has occurred over millennia, then perhaps it would be courteous to publish your babi and other tasks as they are clearly cited in the paper as having been implemented. This will allow the community to verify (and more importantly repeat) your results and build on the ideas within - as you have done so with other giant's work. Otherwise your paper is in effect proof by photoshop/reputation and only a few Nature reviewers will have been able to peer review the content and implementation.
If your intention was to give a public a flavour of your invention and you want to commercialise on this, by all means you and your colleagues at DeepMind are privileged human beings, and you should have not published this as a paper but as a patent and kept us lower mortals away from your contribution.
It seems from many of the issues raised on this project that people want more clear examples of how to use your code. Most of us have waited patiently for 6 months for you to source this code, having been baited in by media hype. Surely as a courtesy to the public it would be nice to give some more concrete examples and documentation as per your paper.
There are others, who are not affiliated to you but have embraced the spirit to the coding community:
https://github.com/Mostafa-Samir/DNC-tensorflow
Forgive my insolence and by no means do I wish to offend you in any way, just wanted to add my opinion.