Open rbharath opened 7 years ago
While the results are cool -- I am failing to see how they transfer to Chemistry problems.
The problem that these networks solve are One Of N classification where new classes are added after training time.
Also I did a cursory pass at attempting to apply this to a generalized multi-task classification/regression problem.
The original matching networks paper was also for one of N w/ new classes added in :-)
I haven't yet taken a close look, but I suspect the attention recurrent comparators can be ported over similarly to how the matching networks and iterative recurrent comparators were. Will try to take a closer look when I get a chance.
I'm marking this a good intermediate contribution for those with deep learning expertise. Also, I suspect this implementation might be easier done with TensorFlow Eager than with graph structured tensorflow (but could be wrong on this hunch).
https://medium.com/@sanyamagarwal/understanding-attentive-recurrent-comparators-ea1b741da5c3
We might want want to add an implentation to dc.metalearning