Optimization of Machine Learning Jobs in the Cloud

João Pedro Neves Nogueira

Informações chave

Autores:

João Pedro Neves Nogueira (João Pedro Neves Nogueira)

Orientadores:

Paolo Romano (Paolo Romano)

Publicado em

28/01/2021

Resumo

Machine Learning and Cloud Computing have been two of the fastest growing areas in the the past few years. Recent developments have emerged regarding machine learning optimizations, enhancing their accuracy and training time. However, for optimization procedures that have very large datasets and many hyperparameters, most users turn to the cloud to offload the inherent computation that would otherwise be infeasible locally. To do so, users face the task of picking cloud parameters to deploy their machine learning jobs which can be difficult due to the wide range of possible configurations and whose misconfiguration translates into large costs for large scale models. Recent state-of-the-art systems have taken the approach of performing optimization of both cloud configurations and machine learning hyperparameters in a joint fashion, with the goal of minimizing cloud related expenses while reaching the best hyperparameter configuration for the specified machine learning algorithm. Nonetheless, the optimization procedure involved by these approaches can also require exploring a large number of expensive configurations and impose, in its turn, large economical costs. It is thus crucial that the optimization procedure is as time efficient as possible and converges rapidly towards the optima. This thesis proposes Hydra, a self-tuning system solution that performs optimization of machine learning algorithms improving some drawbacks of extended state-of-the-art systems by rapidly converging towards the optimum solution without wasting time on bootstrapping the model, using many low-budget evaluations of configurations while applying transfer-learning to enhance the models’ performance, ultimately reducing overall costs by 35% of the extended work.