Basic neural network subroutines (BNNS)
BNNS is a submodule of Accelerate, containing convolution NN primitives optimized for running inference on CPU. It was introduced in iOS 10 and macOS 10.12. Note that it contains only functions for inference, not for training.
The motivation behind this library was to provide unified API for common routines, such that app developers wouldn't need to re-implement convolutions and other primitives from scratch every time (which is hard, as we have seen already in the chapter on CNNs). In a typical CNN, most energy is spent in the convolution layers. Fully connected layers are more expensive computationally, but usually CNNs contain one or a few of them at the very end so convolutions still consume about 70% of energy. That's why it's important to have highly-optimized convolution layers. Unlike MPS, CNN is available on iOS, macOS, tvOS, and even watchOS. So, if you want to run deep learning on a TV set or your watch (just because you can), this is your...
