Uncertainty Quantification of ML Models by L. A. Æ. Sluijterman (.PDF)
File Size: 33.4 MB
Uncertainty Quantification of Machine Learning Models by L. A. Æ. Sluijterman
Requirements: .ePUB, .PDF reader, 33.4 MB
Overview: Machine Learning models have become significantly more popular in recent years and are increasingly being used in areas where reliability is crucial. Think, for example, of self-driving cars or analyzing CT scans. To have confidence in a model, it is necessary to quantify its uncertainty. Since Machine Learning models differ from classical models in crucial ways – they typically have more parameters than data points and are slightly different each time they are created – classical techniques for quantifying uncertainty cannot be directly applied. This work provides new contributions to address this problem. Machine Learning has seen an enormous rise over the past decades. Due to the exponential growth in computing power, machine-learning models have evolved from basic neural networks and decision trees, capable of performing straightforward tasks, to vastly complex architectures that may have billions of parameters. As the capabilities of Machine Learning grow, so does its integration into safety-critical applications such as medical-image analysis, self-driving cars, and the prediction of natural disasters. For these applications, it is essential that these models are trustworthy. A trustworthy model requires trustworthy uncertainty estimates. Relying merely on predictive performance is insufficient. However, producing these uncertainty estimates is far from trivial. Modern models can easily have millions of parameters, making the direct use of many classical techniques difficult or outright impossible. It is this problem of developing new methods to quantify the uncertainty in the predictions of machine-learning models that this thesis contributes to.
Genre: Non-Fiction > Tech & Devices
Free Download links: