Deep Learning Uncertainty using Self-Supervised Learning
Title: |
Deep Learning Uncertainty using Self-Supervised Learning |
DNr: |
Berzelius-2021-80 |
Project Type: |
LiU Berzelius |
Principal Investigator: |
Per Sidén <per.siden@liu.se> |
Affiliation: |
Linköpings universitet |
Duration: |
2021-12-01 – 2022-06-01 |
Classification: |
10106 |
Keywords: |
|
Abstract
Deep learning uncertainty has been shown to greatly benefit from distance awareness. Spectral-normalized Neural Gaussian Process (SNGP, [1]) demonstrated that a deep neural network with distance awareness through spectral normalization combined with a distance-aware Gaussian process output layer can lead to high-quality uncertainty estimates using a single deterministic network. This could have large implications for safety-critical real-time applications in need of robust uncertainty predictions such as e.g. autonomous vehicles. However, the theoretical foundation for distance-awareness via spectral normalization has some flaws, and in this project we instead propose to investigate the potential of achieving distance awareness through self-supervised learning methods.
[1] Simple and Principled Uncertainty Estimation with Deterministic Deep Learning via Distance Awareness (https://arxiv.org/abs/2006.10108)