11–26 Nov 2021
Europe/Budapest timezone

Playing detective: Dissecting silent failures of Deep Learning models for 3D Point Clouds

Not scheduled
20m
Online lecture

Speaker

Robert-Zsolt Kabai (Continental)

Description

Silent failure effects in Deep Learning models are challenging, to say the least. Image processing applications have developed various tools in order to help explainability and debugging, including visualizign kernel activations, heat maps, and so on. Learning on 3D point clouds is especially challenging as we deal with an unstructured data of a point set defined in 3D metric space. Models operating on this native format face multiple challenges compared to models working on images, including rotation and permutation invariance.
This talk will show an example of a silent failure in a 3D point cloud model and the detective work done in order to decode and understand its inner workings. Following a rigorous approach helps us draw the right conclusions as well as aid us in deciding where to start the re-design.

Title

Team Lead, Deep Learning for Point Cloud

affiliation Continental
authors Robert-Zsolt Kabai

Primary author

Robert-Zsolt Kabai (Continental)

Presentation materials