AutonoML: Towards an Integrated Framework for Autonomous Machine Learning.

摘要: Over the last decade, long-running endeavour to automate high-level processes in machine learning (ML) has risen mainstream prominence, stimulated by advances optimisation techniques and their impact on selecting ML models/algorithms. Central this drive is appeal of engineering a computational system that both discovers deploys high-performance solutions arbitrary problems with minimal human interaction. Beyond this, an even loftier goal pursuit autonomy, which describes capability independently adjust solution over lifetime changing contexts. However, these ambitions are unlikely be achieved robust manner without broader synthesis various mechanisms theoretical frameworks, which, at present time, remain scattered across numerous research threads. Accordingly, review seeks motivate more expansive perspective what constitutes automated/autonomous system, alongside consideration how best consolidate those elements. In doing so, we survey developments following areas: hyperparameter optimisation, multi-component models, neural architecture search, automated feature engineering, meta-learning, multi-level ensembling, dynamic adaptation, multi-objective evaluation, resource constraints, flexible user involvement, principles generalisation. We also develop conceptual framework throughout review, augmented each topic, illustrate one possible way fusing into autonomous system. Ultimately, conclude notion architectural integration deserves discussion, field risks stifling its technical advantages general uptake.