Using machine learning to decide what structure to use for your neural net: https://arxiv.org/pdf/1703.01041.pdf
It's pretty meta, but ML is often good at trading computation for results.
miniblog.
Related Posts
Generalised computation is becoming less common: it's much easier to run or debug arbitrary code on a Linux/BSD machine than anything else.
It's even worse on mobile though: simply *writing* is hard and creation is hindered further.
Fun article on refactoring J to be point free, and drawing trees to model its computation:
A new 16-bit floating point format for machine learning! https://hub.packtpub.com/why-intel-is-betting-on-bfloat16-to-be-a-game-changer-for-deep-learning-training-hint-range-trumps-precision/
It increases range at the expense of precision, and often allows 16-bit computation (smaller, faster hardware) to replace 32-bit ML logic.
