p. 261: "weights_error" variable is not being used explicitly

[kovla]

I really appreciate the book, which is well focused on the underpinnings of deep learning and gets the reader up to speed very efficiently. A small detail, in the following code the variable weights_error seems to be redundant, its calculation is never used:

def backward(self, output_error):
    input_error = np.dot(output_error, self.weights.T)
    weights_error = np.dot(self.input.T, output_error)

    #  accumulate the error over the minibatch
    self.delta_w += np.dot(self.input.T, output_error)
    self.delta_b += output_error
    self.passes += 1
    return input_error

Instead, the same calculation is repeated for the accumulation of self.delta_w a couple of lines further. Computational inefficiency might easily be ignored in the educational setting. At least in my own case, however, linking the code to specific mathematical expressions outlined in the book, became less apparent.

One might want to re-write the code as self.delta_w += weights_error, for making the link (e.g. to eq. 10.13) more explicit. I imagine deviating from the book is undesirable at this point, so perhaps a comment could be added to the code on github to make the code flow more explicit?

[rkneusel9]

Thanks for pointing that out. I made the change to avoid the double calculation.