Successive Halving Top-k Operator

Michał Pietruszka, Łukasz Borchmann, Filip Graliński

Abstract: We propose a differentiable successive halving method of relaxing the top-k operator, rendering gradient-based optimization possible. The need to perform softmax iteratively on the entire vector of scores is avoided by using a tournament-style selection. As a result, a much better approximation of top-k with lower computational cost is achieved compared to the previous approach.