Towards meta-learning based, domain specifc AutoML systems in the example domain of cellular image analyses

Roberg, Jan-Henner (2023) Towards meta-learning based, domain specifc AutoML systems in the example domain of cellular image analyses. Master's Thesis / Essay, Artificial Intelligence.

Preview	Text Master_Thesis_Roberg.pdf Download (29MB) | Preview
	Text toestemming.pdf Restricted to Registered users only Download (124kB)

Abstract

Biomedical analyses of cellular images is a time consuming and costly process. For many analyses tasks, semantic segmentation is the first step. Deep Learning has shown great promise in automating this task and is used in existing tools for cellular image analyses. These tools are being used by biomedical experts and thereby facilitate cellular research. However, they rely on default models for the segmentation task and do not tune hyperparameters on a given dataset. This can result in suboptimal performance. In this thesis we propose an automated machine learning (AutoML) system to tune the hyperparameters of a DL-pipeline for semantic segmentation of cellular images. The system is based on meta-learning and leverages meta data to efficiently find a suitable pipeline for a new dataset. Meta data here refers to DL-pipelines evaluated on different datasets. To quantify datasets, different meta-features are extracted. These meta-features are concatenated with hyperparameter configurations and used as input to a machine learning model. This model is trained on the meta base and learns to predict how well hyperparameter configurations will perform on different datasets. After training, this model can then predict a ranking of hyperparameter configurations for a new dataset. Based on the predicted ranking successive halving is used to determine the best hyperparameter configuration for a dataset. In addition to the AutoML system, the effects of generic and domain specific transfer

Item Type:	Thesis (Master's Thesis / Essay)
Supervisor name:	Jaeger, H.
Degree programme:	Artificial Intelligence
Thesis type:	Master's Thesis / Essay
Language:	English
Date Deposited:	30 Jan 2023 10:56
Last Modified:	30 Jan 2023 10:56
URI:	https://fse.studenttheses.ub.rug.nl/id/eprint/29178

Actions (login required)

View Item