SUPR
Machine Learning for Visual Perception, storage
Dnr:

NAISS 2024/6-300

Type:

NAISS Medium Storage

Principal Investigator:

Per-Erik Forssén

Affiliation:

Linköpings universitet

Start Date:

2024-09-27

End Date:

2025-10-01

Primary Classification:

10207: Computer Vision and Robotics (Autonomous Systems)

Secondary Classification:

10105: Computational Mathematics

Tertiary Classification:

10106: Probability Theory and Statistics

Allocation

Abstract

This project studies information representation for visual perception. As can be inferred from the numerous examples of visual illusions, visual perception is often ambiguous. This makes it important that perceptual inference is coupled with statements of uncertainty, that quantify whether the percept is ambiguous, or if there are several potentially valid explanations of the sensory data. Research is conducted by designing deep neural networks that take images, video, or point-cloud data as input. The networks predict percepts, and their associated uncertainties, using representations such as probability density functions, and ensembles. A major focus is the design of appropriate objective functions that facilitate learning while avoiding problems of overconfidence.