patgpt4's picture
Duplicate from Anonymous-sub/Rerender
1239b39
|
raw
history blame
670 Bytes

FAQs

Q: If the weight of a conv layer is zero, the gradient will also be zero, and the network will not learn anything. Why "zero convolution" works?

A: This is wrong. Let us consider a very simple

y=wx+by=wx+b

and we have

y/w=x,y/x=w,y/b=1\partial y/\partial w=x, \partial y/\partial x=w, \partial y/\partial b=1

and if $w=0$ and $x \neq 0$, then

y/w0,y/x=0,y/b0\partial y/\partial w \neq 0, \partial y/\partial x=0, \partial y/\partial b\neq 0

which means as long as $x \neq 0$, one gradient descent iteration will make $w$ non-zero. Then

y/x0\partial y/\partial x\neq 0

so that the zero convolutions will progressively become a common conv layer with non-zero weights.