Cloud AI needs a data centre. Edge AI runs on a battery. Quantise weights to a quarter of the size, prune away most of a network, and budget energy like it costs money.
3 units, about 3 hours. Free.
Quantise weights to 8-bit integers, measure the size saving and the accuracy cost.
2Prune weights by magnitude, measure the accuracy trade, and explain structured versus unstructured pruning.
3Budget energy for a battery device and explain why radio use dominates every other cost.