多任务/多场景学习技术分享

综述

 

Grad Norm

梯度归一化

https://mp.weixin.qq.com/s/RIxxtMqdb6yJKLorg_WjrA

https://www.cnblogs.com/douzujun/p/14633524.html

多任务学习的多个权重改如何调节

梯度修剪，保证每个任务的梯度在同一个数量级

https://blog.csdn.net/Leon_winter/article/details/105014677

 

多场景论文

DADNN: Multi-Scene CTR Prediction via Domain-Aware Deep Neural Network

One Model to Serve All: Star Topology Adaptive Recommender for Multi-Domain CTR Prediction

https://zhuanlan.zhihu.com/p/361113529

摘要

 

ple

https://zhuanlan.zhihu.com/p/291406172

代码

摘要