Validation of Uncertainty in Classification under Point Cloud Downsampling

Khalil, Jay (2023) Validation of Uncertainty in Classification under Point Cloud Downsampling. Bachelor's Thesis, Artificial Intelligence.

Preview	Text Thesis_s3167240.pdf Download (4MB) | Preview
	Text toestemming.pdf Restricted to Registered users only Download (129kB)

Abstract

The rapid evolution of artificial intelligence (AI) highlights the need to understand uncertainties in machine learning (ML) predictions. This research explored uncertainties in point cloud data, a high-dimensional structure pivotal in many fields, utilising Bayesian neural networks. Hence, we integrated Monte Carlo Dropout, Monte Carlo DropConnect, Flipout, and Deep Ensemble within a custom PointNet model. Through iterative downsampling, the models unveiled varying abilities to manage uncertainties. Our findings suggest a prevalent underconfidence due to point cloud data complexity. Importantly, model sensitivities varied across classes, indicating the intricacies of class representation and training complexities. Notably, the extent of downsampling influenced uncertainty in non-linear ways, challenging the assumption that more sampling points always yield better results. These insights underscore the study's contribution to ML's uncertainty management, hinting at avenues for model optimisation. Future research should delve deeper into this promising domain.

Item Type:	Thesis (Bachelor's Thesis)
Supervisor name:	Valdenegro Toro, M.A.
Degree programme:	Artificial Intelligence
Thesis type:	Bachelor's Thesis
Language:	English
Date Deposited:	16 Aug 2023 09:29
Last Modified:	16 Aug 2023 09:29
URI:	https://fse.studenttheses.ub.rug.nl/id/eprint/31190

Actions (login required)

View Item