University thesis:
Dissertation, Universität Freiburg, 2023
Footnote:
Description:
Abstract: Deep learning models require significant resources to deploy, limiting their widespread adoption. The aim of this thesis is to address this issue by proposing methods to make deep learning models more efficient for training and deployment.<br><br>One important aspect of machine learning is the ability to understand visual information from limited labeled data because large-scale annotation processes can be very expensive or infeasible. The first part of the thesis proposes methods to improve label efficiency for deep learning-based computer vision tasks focusing on two approaches - semi-supervised learning and active learning. For semi-supervised learning, the thesis proposes an approach for semi-supervised semantic segmentation that learns from limited pixel-wise annotated samples while exploiting additional annotation-free images. The proposed dual-branch approach reduces both the low-level and high-level artifacts typically encountered when training with few labels, and its effectiveness is demonstrated on several standard benchmarks. For active learning, the thesis emphasizes that conventional evaluation schemes used in deep active learning are either incomplete or below par. The thesis investigates several existing methods across many dimensions and finds that the studied new underlying factors are decisive in selecting the best active learning approach. The thesis also provides a comprehensive usage guide to obtain the best performance for each case. This thesis covers active learning methods for image classification and semantic segmentation tasks.<br><br>Another issue with deep neural networks is catastrophic forgetting when encountering new or evolving tasks in a sequential manner. The model must be retrained with all the data or tasks encountered to avoid forgetting, thus making them unsuitable for many real-world applications. The second part of the thesis focuses on understanding and resolving catastrophic forgetting in continual learning, particularly in the Class-incremental Learning (CIL) setting. The evaluation shows that a combination of simple components can already resolve catastrophic forgetting to the same extent as more complex measures proposed in the literature.<br><br>Overall, this thesis provides streamlined approaches to improve the efficiency of deep learning systems and highlights the importance of many unexplored directions for improved realistic evaluation