| dc.description.abstract | Machine learning has succeeded in real-world applications from image classification, speech recognition, to beating human champion in Go games. To accelerate the development of different ap-plications, automated machine learning (AutoML) has been proposed to discover high-performance machine learning models automatically. It could release the burden of data scientists from the multifarious manual tuning process. However, dataset does not always have correct labels and sufficient data size. Wrong labels disrupt the training procedure and could not provide representative evaluation performance for AutoML. Imbalanced label distribution further skews search feedback for AutoML. Furthermore, additional performance constraints, such as model size, fairness, and robustness complicates the AutoML flow. When computing resources are insufficient, small search space constrains the flexibility to search neural networks, which causes inconsistent architectures used in search and evaluation stages. In this dissertation, I advanced AutoML from aspects of imperfect data and constraints. A new robust loss function is integrated with search algorithm for label noise. I design a simple but effective search space for imbalanced defect datasets. The defect generator can alleviate imbalanced distributions. I also proposed constraint-aware early stopping for AutoML with adaptive constraint evaluation intervals. An efficient model parallelism for AutoML is proposed to extend search spaces in multiple GPUs with limited memory size. My research of automated machine learning enables scientists to obtain off-to-shelf models on various data formats, as well as customizes models for different computing resources, model size requirements, and miscellaneous performance constraints. It broadly impacts image classification, constrained AutoML, and defect detection. | |
