Machine learning has attracted tremendous attention from researchers in various fields. In past decades, machine-learning techniques make remarkable progress and get great success across a variety of domains, such as robotics, computer vision, astronomy, biology, etc. One of the things that makes them so fascinating is that they often interact directly with the external world. However, the external world is rarely stable. Applying machine-learning techniques in critical applications, like autonomous driving, requires not only point predictions but also reliable uncertainty measurements. The source of uncertainties in machine learning can be generally classified into three categories: data source, model parameters, and model structures.

In this thesis, we take probabilistic theory as the...[ Read more ]

Machine learning has attracted tremendous attention from researchers in various fields. In past decades, machine-learning techniques make remarkable progress and get great success across a variety of domains, such as robotics, computer vision, astronomy, biology, etc. One of the things that makes them so fascinating is that they often interact directly with the external world. However, the external world is rarely stable. Applying machine-learning techniques in critical applications, like autonomous driving, requires not only point predictions but also reliable uncertainty measurements. The source of uncertainties in machine learning can be generally classified into three categories: data source, model parameters, and model structures.

In this thesis, we take probabilistic theory as theoretical foundations and adopt error propagation as well as Laplace approximation to estimate the uncertainty in various stages of a perceptual algorithm. We take LiDAR-based 3D object detection as an instance to demonstrate the feasibility and apply it in the context of autonomous driving. For input point clouds, we model the uncertainties in extrinsic parameters of a multi-homogeneous-LiDAR system and propagate them into each point to improve the robustness of algorithms in geometric tasks. In weight space, we evaluate the posterior distribution of weight parameters in deep neural networks with Laplace approximation and adopt it as an uncertainty measurement for each parameter. They are further used to compute Bayesian constraints for preserving old-task knowledge along with knowledge distillation regularizers. In predictions, we tailor Laplace approximation methods to estimate epistemic uncertainty for 3D object detection. Compared to conventional point-estimation perception results, the prediction with epistemic uncertainty enriches the result representation and provides the reliability information for downstream tasks. In conclusion, this thesis demonstrates the feasibility of uncertainty estimation in various stages of deep neural networks and its application in multi-sensor fusion, incremental learning, and reliable predictions. Some promising future works are also discussed at the end of this thesis.