论文标题
连续减半TOP-K操作员
Successive Halving Top-k Operator
论文作者
论文摘要
我们提出了一种放松顶级运算符的连续减半方法，从而使基于梯度的优化成为可能。通过使用锦标赛风格的选择，可以避免在整个分数向量上进行软智能迭代。结果，与以前的方法相比，TOP-K的近似值要较低。
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.
