Session: 12-11: Software-Defined Vehicles, Digital Twins, and Connected AI Platforms

Paper Number: 197489

197489 - From Universal Approximation to Data Loops: Data‑centric Ai for Safer and More Efficient Connected Fleets 

Abstract: 

Neural networks are often introduced through the universal approximation theorem: with enough capacity, a network can approximate any continuous function to arbitrary accuracy. This result helps explain why deep learning is so powerful, but it can also be misleading in practice. In real‑world mobility systems, we rarely lack model capacity; instead, we lack the right data, in the right regions of the input space, with the right labels. The limiting factor is not whether a function can be approximated, but whether we have the data needed to approximate it where it matters most for safety and energy efficiency.

This talk connects the universal approximation perspective to a practical data loop for connected fleets: deploy a model, collect data, automatically create or refine labels, retrain, and redeploy—repeating this cycle to progressively improve performance. I will describe a three‑part “3A” framework:

Active learning: use ensembles or variants of a model to identify inputs where predictions disagree or uncertainty is high, marking them as “interesting” for further attention.

Auto‑labeling: use offline perception models and vision‑language models to propose labels for these interesting samples without requiring humans in the loop for every example.

Augmentation: use generative techniques to create synthetic variations of rare but important scenarios (for example, modifying scenes, signs, or traffic light states) to expand coverage of safety‑critical corner cases.

Presenting Author: Dalong Li Samsara

Presenting Author Biography: Dalong Li is a leading expert in computer vision and machine learning with over 20 years of experience spanning foundational AI research and large-scale industrial deployment in autonomous systems. He currently serves as a Senior Applied Scientist at Samsara, where he leads research and development in Perception and AI Safety, focusing on the next generation of connected operations and vision-based safety systems.
Prior to joining Samsara, Dr. Li was the Head of Auto Labeling and Data Loop at Torc Robotics (a subsidiary of Daimler Truck), where he architected and scaled the AAA Framework (Auto Curation, Auto Labeling, and Augmentation). His work was instrumental in transforming offline perception pipelines, utilizing Knowledge Distillation and Active Learning to solve complex "long-tail" scenario mining challenges for Level 4 autonomous trucking.
A recognized voice in the field of technical integrity and safety-critical AI, Dr. Li is a co-author of international industry standards, including ISO/TR 4804 (Safety and Cybersecurity for Automated Driving Systems). His career bridges deep academic rigor with real-world engineering, having spent decades developing robust perception systems for diverse and challenging environments. Dr. Li received his Ph.D. in Electrical and Computer Engineering from the Georgia Institute of Technology (Georgia Tech). He is a frequent speaker at major industry summits, including the IEEE AI Southeast Michigan Summit, and is a committed advocate for Data-Centric AI and sensing safety.

Authors:

Dalong Li Samsara