This paper/link talks about resnet's bottleneck design.

enter image description here It's totally not clear to me how the bottleneck design on the right is equivalent to the left-diagram and how is it reducing the parameters? Left has 64-d input whereas right has 256-d. What is 'd' denote? If it's channel in the input, then what does 64 in '1x1, 64' denote? If the input to 1x1 conv is 256 channel, then doesn't it implicity mean, we have a filter of 1x1x256 for the conv? So not sure, where that 64 comes from.

In short it's very unclear how left-side is equivalent to the bottneck diagram on right and how right is better. Could someone please explain?


1 Answer 1


They're not equivalent, the claim is just that the block on the right has more expressivity while being more computationally efficient.

Number of parameters of each layer can be computed as $NMK^2$ where the kernel is $K \times K$ and there are $N$ input and $M$ output channels. So the left block has about 74k parameters and the right block has 70k.

d denotes the channels in the output of each layer.

  • $\begingroup$ I see. Any reason then why the paper/link above compares these two in the diagram and not the ones that would have been equivalent? $\endgroup$
    – Joe Black
    Apr 13, 2020 at 18:35

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.