Some changes to the script to produce the plot we use in scipy ( basically fix the subplot problem, and play a bit with the axis) '

added a new optimization for tensor of pow. Disabled by default as this cause slow down on gpu in the current form. To enable it use the flag experimental.pow.

dy*sm - DimShuffle{0,'x'}(sum{1}(dy*sm))*sm -> softmax_grad(dy,sm)

tensor.Canonize - adding external simplifiers to permit adding optimizations to the local_canonicalizers / canonizers