Abstraction, mimesis and the evolution of deep learning

Abstract Deep learning developers typically rely on deep software frameworks (DLSFs)—simply described as pre-packaged libraries of programming components that provide high-level access to functionality. New DLSFs progressively encapsulate mathematical, statistical and computational complexity. Such higher levels abstraction subsequently make it easier for methodology spread through mimesis (i.e., imitation models perceived successful). In this study, we quantify increase in discuss its implications. Analyzing publicly available code from Github, found the introduction correlates both with significant increases number projects substantial reductions lines used. We argue importance respect ephemeralization, technological advancement, democratization, adopting timely abstraction, emergence mimetic deadlocks, issues related use black box methods including privacy fairness, concentration power. Finally, also a symptom an ongoing metatransition.