BriefGPT.xyz
Dec, 2020
PiRank: 可扩展的可微排序学习
PiRank: Learning To Rank via Differentiable Sorting
HTML
PDF
Robin Swezey, Aditya Grover, Bruno Charron, Stefano Ermon
TL;DR
提出了一种新的可微的代理损失函数PiRank,它使用基于NeuralSort的连续、温度控制放松来排序操作符,最终在公共互联网规模的学习排序基准测试中明显提高了可比较方法的性能。
Abstract
A key challenge with
machine learning
approaches for
ranking
is the gap between the performance metrics of interest and the
surrogate loss functi
→