Neural network architecture search for extreme classification and in a partially labeled learning context
Loïc Pauletto · HAL (Le Centre pour la Communication Scientifique Directe) · 2022
Deep learning applications are rapidly expanding and show no signs of slowing down. Neural network topologies are becoming larger and more complex for challenging real-life problems.This increased complexity necessitates more time and expertise from professionals, as well as a significant financial investment for AI companies.Neural Architecture Search is a novel Machine Learning paradigm that seeks to determine the best NN architecture for a given problem. NAS techniques, on the other hand, have only been studied and developed in limited, well-defined Machine Learning problems, which are not representative of all existing ML scenarios.This thesis focuses on the research and development of the NAS approaches for new tasks as well as a new learning framework that is more relevant to real-world applications.We suggested using a neuro-evolutionary NAS framework to solve the extreme multi-label classification challenge in particular.We combined convolution and recurrent networks to provide a more appropriate space search for this assignment.On several datasets, we evaluate the performance of the searched network. We also looked at the challenge of reconstructing an RSSI map, which is a more difficult process due to the lack of input data and the fact that it is only partially annotated. In this way, we provide a system for semantic segmentation task dynamic architecture search with a minimal number of annotated samples. We investigated multiple semi-supervised learning algorithms in this framework to see which one was the most successful at using unlabeled samples.We looked at a number of strategies, including "traditional" and "new" semi-supervision approaches, as well as self-supervision approaches.