Visualizing Convolutional Networks: Max-pooling Switches
Intermediate
The problem, from first principles
To see what a max-pooling layer discards, Zeiler and Fergus recorded the position of each pooled maximum, its switch. Those switches let a deconvolution network place activations back where they came from.
From theory to code
Implement max_pool_switches(x, size), returning the offset of the maximum within each pooling window.
Constraints
- Work on a 1D signal for simplicity.
Hints
Hint 1
Reshape the signal into rows of length size, then take the argmax of each row.
Keywords
research-paperscomputer-visioncnninterpretability
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.