meta-learning experiments

I believe there is a bug at "eval2.py" file. RMSprop run (lines 50-60) is done updating the embedding weights after one iteration, and the converged results is just +- the learning rate.  I believe it happens due to  the embedding weights initialization to zero (lines 20-21). Comment these lines out gives much better loss.