A comparative analysis between Google "Cloud AutoML" and Sagacify in-house deep learning models

Details

Supervisors

Saerens, Marco ; Courtain, Sylvain

Faculty

Louvain School of Management

Degree label

Master [120] en ingénieur de gestion

Abstract

This thesis contains a performance comparative analysis of deep learning algorithms. More precisely, the deep learning performances of the service "Cloud AutoML" from Google Cloud Platform will be compared with the ones of bare metal deep learning algorithms, implemented within the company Sagacify. The goal of this thesis is to determine which model, between the fully custom and the black box, should be preferred for an image classification case, depending on the characteristics of the data set. Here, 6 different characteristics are studied. The assessment is done based on performance metrics (accuracy, precision, recall and F1 score), on the training time, on the related costs, on the explainability and on the accessibility of each solution. After a brief introduction, the thesis will first provide some of the main business applications of artificial intelligence. This will be done in order to highlight the importance of the opportunities that one can capture thanks to this technology. In a second stage, some theoretical concepts about artificial intelligence will be explained. More precisely, machine learning and its subdivisions (supervised learning, unsupervised learning and reinforcement learning) are topics that will be discussed, followed by logistic regression and deep learning. For the latter, deeper explanations will be provided for artificial neural networks, convolutional neural networks and transfer learning. Thirdly, the two entities where the research will be done will also be introduced. Concepts related to cloud services will be briefly explained, as well as the different services provided. Nevertheless, the focus will remain on the cloud service used in this thesis, which is Google Cloud Platform. Then, a high level presentation of the company Sagacify will be given, with a focus on its environment, its services and products, as well as some of its projects. After this introduction of Sagacify, the experimental protocol will be given. More precisely, each step will be explained, which are the data description, the data split, the model construction and selection, the model training and the comparison methodology. After the experimental protocol, the following section will discuss and analyse the results of the research. There, the performance metrics will be compared, as well as other relevant aspects. Based on them, their business implications for Sagacify will also be discussed. Finally, the thesis will end with a conclusion, highlighting the major outcomes of the research, but also its limits, as well as providing advices and directions for further researches.