Approach

Fixed-point Factorization

SDD

Full-precision Weights Recovery

The quantized weights are used to conduct forward and backward computation and the full precision weights are used to accumulate gradients during back-propagation.

Experiment

References:
Fixed-point Factorized Networks, Peisong Wang, 2017, CVPR

Fixed-point Factorized Networks
Approach Fixed-point Factorization Full-precision Weights...
文章点评CV9-Factorized Inference in
0. 基本信息题目：Factorized Inference in Deep Markov Models for...
Factorized Convolutional Neural
摘要深度卷积神经网络在视觉识别领域已经到达非常客观的性能，但这些都是在高计算复杂度的代价下实现的。这篇文章提出了...
Factorized Convolutional Neural
Approach By unravelling channel and spatial convolution a...
Graph Neural Networks: A Review
Graph Neural Networks Graph Neural Networks: A Review of ...
第一代图卷积网络：图的频域网络与深度局部连接网络
论文标题：Spectral Networks and Locally Connected Networks on ...
LSTM
Recurrent Neural Networks networks with loops in them, al...
Fixed-point Quantization of Conv
这篇论文主要使用位宽（Bitwidth）和小数偏移（Fractional Offset）这两个参数来表征定点数，其...
05组——Non-local Neural Networks
Non-local Neural Networks Non-local Neural Networks是何凯明大佬...
Identity Mappings in Deep Residu
Residual-networks可以看作是highway-networks的特殊版本，特别是“Identity ...