Maxout Networks: Which Piece Wins?
Intermediate
The problem, from first principles
The maxout unit only sends gradient through the piece that wins. Knowing which piece wins for each input is what turns the max into a piecewise-linear function, and it determines the backward pass.
From theory to code
Implement maxout_argmax(x, W, b, k), which returns the index of the winning piece for each unit and example.
Constraints
- Indices are in
[0, k-1].
Hints
Hint 1
Use the same grouping as maxout_forward, but call argmax on the last axis instead of max.
Keywords
research-papersactivationmaxout
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.