Gustavo Rodrigues dos Reis

Abstract: 
Deep learning (DL) models are powerful tools for analyzing data, but with so many models available, it can be hard to know which one is best for a specific use case. This thesis work adapts processes from the Case-Based Reasoning (CBR) methodology, which seeks solutions to new problems by finding similar problems that have been solved in the past. In this thesis context, a DL model is seen as a solution to a specific problem. The designed approach then leverages metadata descriptions of DL models and datasets to help users identify pertinent deep learning models for their use case, amongst existing models. Given the rapidly evolving landscape, the proposed system design prioritized modularity and agile orchestration. Modularity is required to integrate new model structures, datasets, and comparison algorithms as they are developed, and agile orchestration when isolating software components, allowing models with different execution requirements to be exchanged more efficiently. Components of the approach were therefore prototyped within a hybrid no-code/low-code platform within the industrial collaboration with NAVER LABS Europe.

Keywords : 
Case-Based Reasoning, Deep Learning, Software Artifact Reuse, Pipeline Agility, Component Modularity, Curriculum Learning, Hyperparameter Optimization