Over the past decades we have normally never bothered with the exact needs of an implementation when it comes to computing accuracy (double FP, FP, Integer-64 / -32 etc.). Nevertheless there are important consequences to energy use depending on the chosen accuracy. This area of 'approximate computing' is also important from a mathematical point of view.  Given our future of big-data and Internet of Things (IoT), quasi-real time or even real-time processing of data into information will become essential, and choosing the appropriate accuracy and having appropriate parallel methods (links back to the previous point) will be crucial to achieve this. One might think of a neural network to be the ultimate approximate-computing optimized massively-parallel way to address IoT processing?