Preparing Active Learning for the real world

Julius Gonsior

AutorIn

Julius Gonsior

Titel

Preparing Active Learning for the real world

Zitierfähige Url:

https://nbn-resolving.org/urn:nbn:de:bsz:14-qucosa2-1004835

Erstveröffentlichung

2025

Datum der Einreichung

05.06.2025

Datum der Verteidigung

23.10.2025

Abstract (EN)

Machine learning is a standard artificial intelligence method that enables computers to learn patterns from data and make predictions or decisions without explicit instructions. Most machine learning problems rely on labeled datasets containing training examples and their desired output as labels. Often, the labels for a concrete use case are missing and need to be supplied somehow, usually by human annotators. This is called the data labeling bottleneck. As the process of labeling a dataset is a time-consuming and expensive task, this often prevents machine learning projects from being successful. To overcome this bottleneck, there are several computational methods that significantly reduce the required labeling effort. Among them, active learning is universally applicable without prerequisites. Still, despite being known for decades, it is not a widely adopted method. The primary goal of this thesis is to investigate what prevents active learning from widespread usage by practitioners and propose solutions to overcome the found obstacles. Our research focuses on four main areas: We a) design a combination of active learning and weak supervision, an alternative computational method for overcoming the data labeling bottleneck, b) propose an adaptive active learning strategy that is trained on synthetic datasets, freeing practitioners from the need to choose a specific strategy, c) present an easy-to-implement solution to ignore harmful outliers during the active learning process in combination with transformer-based neural networks, and d) create and conduct a benchmark which investigates all significant active learning hyperparameters and makes active learning evaluations more reproducible. Our individual achievements make active learning more accessible for practitioners. Therefore, this will allow more unlabeled datasets to be used for real-world use cases and, ultimately, allow more machine learning projects to become a reality.

Freie Schlagwörter (EN)

Active Learning, NLP, Machine Learning

Klassifikation (DDC)

006

Klassifikation (RVK)

ST 306

GutachterIn