Network in Network: The 1x1 Convolution
Intermediate
The problem, from first principles
Network in Network (Lin, Chen & Yan, 2013) replaces a convolution's linear filter with a small multilayer network applied at every position. A 1x1 convolution is the simplest form: a dense layer over channels, reused at each pixel.
From theory to code
Implement conv1x1(x, W, b), which maps an (H, W, C) feature map to (H, W, C_out) with one linear layer per pixel.
Constraints
- The same
Wandbare used at every spatial position.
Hints
Hint 1
Apply x @ W.T + b; NumPy broadcasts the matrix multiply over the first two axes.
Keywords
research-papersarchitectureconvolution
Python
Run your code to see results
No results yet
Run your code to check it against the examples, or submit to run every test.