Deep learning works when you follow every step of a course assignment, or a research paper. But if you have tried your own deep learning project or designing a new kind of layer, 99% chance that the model learnt nothing. This post gets to the heart of the one of the many reasons your model might not work, i.e. the gradients are broken.
Part of course:
Andrej Karpathy: Yes you should understand backprop
Show admin stats