I wonder at what stage notions of metabolism will be baked directly into neural network operations -- e.g optimize not just for reducing some notion of error on a dataset, but also for the training efficiency itself.