Engineering framework for scalable machine learning operations

Abstract

With the evolution of algorithms and solutions in the artificial intelligence field, new and modern methods and practices are required to successfully leverage these technologies. Therefore, a new field named Machine Learning Operations (MLOps) has grown rapidly in the last five years. The goal is to increase integration between the pure research in the artificial intelligence domain and the traditional software engineering domain to generate business value faster, by rapidly shifting machine learning algorithms to the production stage.

The thesis aims to provide a novel machine leaning framework to exploit the business potential of solutions to artificial intelligence and data mining problems in the industry by introducing novel tools and practices, which ensure automation and maintenance are guaranteed on the long run. In particular, the thesis introduces the Model-as-a-Package concept. The design considers a machine learning pipeline as a highly specialized package. Based on this, a novel framework is proposed and implemented. The ultimate goal is to enable robust automation in the model versioning, model deployment and monitoring aspects of machine learning operations.

Compared to the state-of-the-art end-to-end open source frameworks in the market, the features the framework introduces outperform the alternatives in the context of model life cycle processes. In particular, major improvements have been proposed for dependency management, automated logging and automatic model update.

Moreover, the framework improves the state of the art by decoupling logic into two highly integrated components, namely Trainer and Predictor. This expands the set of inference services supported by the framework by adding low resources serving channels such as serverless functions and IoT devices.