Bringing neural networks to cellphones

I doubt pruning alone is going to serve up the same kind of wins that energy aware architectures will. The main problem is that convolutions are heavy computational operations. There has been ongoing research in modifications to the convolution that basically decrease the amount of computation being done in channel space (grouped and depthwise separable convolutions). The tradeoff is accuracy, but new techniques (like shuffling the channels) are helping recover that.

MobileNets and ShuffleNet are some examples. They see 10x reductions in MACs, far more than the "73 percent reduction in power consumption over the standard implementation of neural networks" talked about in this article. https://arxiv.org/pdf/1704.04861.pdf https://arxiv.org/pdf/1707.01083.pdf

Designing Energy-Efficient Convolutional Neural Networks using Energy-Aware Pruning: https://arxiv.org/pdf/1611.05128.pdf

There are already providers of building blocks for the fabless SoC industry, that use various NN schemes. For example to better detect signal in noise like in iterative codes [0, 1].

I remember a meeting with such a company, that offered building blocks to classify signals, but unfortunately I do not remember its name.

[0] http://ieeexplore.ieee.org/document/1180574/

[1] http://ieeexplore.ieee.org/document/1618080/