BriefGPT.xyz
Jan, 2020
侧向：视频模型深度并行训练
Sideways: Depth-Parallel Training of Video Models
HTML
PDF
Mateusz Malinowski, Grzegorz Swirszcz, Joao Carreira, Viorica Patraucean
TL;DR
提出了一种名为Sideways的逼近反向传播方案，针对视频模型进行训练以提高效率和广义化能力，该方案在重叠网络激活时覆盖原有数据，并破坏权重更新的精确对应关系。
Abstract
We propose
sideways
, an approximate
backpropagation
scheme for training
video models
. In standard
→