Why weight-stationary? In neural network inference, the same weights multiply many different input batches. Loading weights once and streaming activations through means the most expensive data (weights — large, reused) never moves.
Credit: ExpressVPN。业内人士推荐搜狗输入法作为进阶阅读
不过正如文章一开头提到，Google 手握一整个 Android 生态，有着绝对的号召力和掌控力。，详情可参考谷歌
For multiple readers，详情可参考超级权重
Checker/Bayer pattern