Proj CDeepFuzz Paper Reading: SparseProp: Efficient Sparse Backpropagation for Faster Training of Neural Networks
Abstract
本文：SparseProp
Github: https://github.com/IST-DASLab/sparseprop
Task: a back-propagation algo for sparse training data, a fast vectorized implementation on commodity CPUs
特点： general, can apply to arbitrary sparsity and common layer types; the first support for sparse training on commodity hardware
实验：
效果：
- 验证vectorized implementation加速，能在从头训练或者transfer learning两种场景下
