Thema der Dissertation:
Deep Learning-based Localisation for Autonomous Vehicles
Thema der Disputation:
Multi-Task Learning Using Uncertainty to Weigh Losses for Scene Geometry and Semantics.
Deep Learning-based Localisation for Autonomous Vehicles
Thema der Disputation:
Multi-Task Learning Using Uncertainty to Weigh Losses for Scene Geometry and Semantics.
Abstract: Multi-task learning is the approach to improve the generalization of individual deep learning models by training them in parallel and through a shared representation. Several applications, including autonomous vehicles, which have multiple regression and classification objectives, can benefit from multi-task learning. Nevertheless, the approach performance depends highly on a relative weighting between each task's loss, a tunning manual task that makes multi-task learning not convenient in practice. This talk will present and discuss an approach formulated by Kendall, Alex et al. (2018) to effectively developing multi-task learning by evaluating each objective's homoscedastic uncertainty during training and automatically finding an improved model for the tasks. The model is demonstrated by learning simultaneously per-pixel depth regression, semantic, and instance segmentation from a monocular input image1.
1 Kendall, Alex et al. “Multi-task Learning Using Uncertainty to Weigh Losses for Scene Geometry and Semantics.”
2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition (2018): 7482-7491.
Zeit & Ort
20.11.2020 | 10:00