Ternary weight networks

Approach

We introduce ternary weight networks (TWNs) - neural networks with weights constrained to +1, 0 and -1. The Euclidian distance between full (float or double) precision weights and the ternary weights along with a scaling factor is minimized.

三值网络相比于二值网络，阈值需要求解，在求解阈值时，假设W是服从均匀分布的；另外，三值网络比二值网络多了0值，相当于自动做了剪枝，或者说二值网络稀疏化。

Experiment

References:
Ternary weight networks, Fengfu Li, 2016, CVPR