torch / nn

Other
1.34k stars 967 forks source link

Why isn't the error scaled down during Mul's backprop ? #1285

Open singam-sanjay opened 7 years ago

singam-sanjay commented 7 years ago

https://github.com/torch/nn/blob/872682558c48ee661ebff693aa5a41fcdefa7873/Mul.lua#L29-L33 Why isn't the error caused by the input to the Mul layer not divided by the scaling factor, but multiplied by the same ?

If the input to the layer with the scaling factor of 2 is [1, 2] and the expected output is [4,4], the error would be [-2, 0]. Subtracting the scaled down error ([-1, 0]) from the input would correct the input.

Please clarify.