Deep generative models for Bayesian inference on high-rate sensor data: applications in automotive radar and medical...

Area of Science:

• Artificial Intelligence
• Computer Vision
• Signal Processing

Background:

• Deep generative models (DGMs) excel in natural image processing.
• Bayesian methods leverage DGMs for image restoration tasks (denoising, inpainting, super-resolution).
• Growing interest in DGMs for Bayesian inference on sensory data.

Purpose of the Study:

• Address challenges in applying image-centric DGMs to raw sensory data.
• Develop solutions for high dynamic range (HDR) signals, multi-sensor interference, and high data rates.
• Adapt DGMs for complex, often unknown, physical data-generating processes with non-Gaussian noise.

Main Methods:

• Discusses challenges in adapting DGMs for sensory data.
• Proposes approaches to handle HDR signals and multi-sensor interference.
• Explores real-time processing requirements (latency, throughput).

Main Results:

• Highlights the difficulties of direct DGM application to raw sensor data.
• Identifies the need for specialized methods to handle sensory data characteristics.
• Demonstrates the relevance of DGMs in automotive radar and medical imaging.

Conclusions:

• Adapting DGMs for sensory data requires addressing specific technical hurdles.
• Proposed approaches aim to bridge the gap between DGM capabilities and real-world sensing needs.
• This work contributes to a new paradigm for inverse problems using generative modeling and Bayesian inference.