presented by Piotr Kozakowski
Sorting and ranking are important.
Sorting and ranking are important.
But not differentiable.
Sorting and ranking are important.
But not differentiable.
Sorting: piecewise linear - continuous, derivatives constant, zero or undefined.
Sorting and ranking are important.
But not differentiable.
Sorting: piecewise linear - continuous, derivatives constant, zero or undefined.
Ranking: piecewise constant - discontinuous, derivatives zero or undefined.
Goal: construct differentiable approximations of sorting and ranking.
Convex hull of permutations on w.
source: the paper
source: Wikipedia
same solutions
from the Fundamental Theorem of Linear Programming
Euclidean projection of z onto the permutahedron!
(not going to talk about that)
TLDR: we can pose the problem as isotonic regression
and solve it in O(n log n) time and O(n) space.
And we can multiply with the Jacobian in O(n).
(it's sparse)
Idea: sort the losses and ignore the top k of them.