Labs

Convolution

Assume 1) Locality 2) Stationary

convolution is for computation efficiency based on the above two assumption. If we shuffle the input, it loses locality and stationary, we cannot use convolution.

20 Feb

Batch Normalization

https://gab41.lab41.org/batch-normalization-what-the-hey-d480039a9e3b

Larger batch size, Matrix Matrix computation can be optimized. (so, AlexNet feedforward time drops when batch size increase)