Data selection is crucial for creating diverse datasets, enabling active learning, detecting corner cases, and building new AI functions.
Data can be sourced from data lakes with heavy compute, targeted test fleets, or the vast customer fleet which offers real-world scenarios but has limited compute.
Informative data lies in the long tail of the data distribution, including rare scenarios, weak sensor signals, and atypical class distributions.
Key methods for selecting valuable data include uncertainty estimation, temporal analysis of predictions, anomaly detection, and using model ensembles.
An uncertainty trigger aggregates softmax scores from a traffic light detection model to identify and record challenging images like false positives or distant objects.
Proper model evaluation requires testing against not just random data but also corner-case datasets to prevent performance regressions in specific scenarios.
An in-vehicle module called "Instinct" filters data streams in real-time, enabling continual learning by collecting data from new regions to expand a model's operational domain.
A universal framework uses a plugin architecture to support various trigger types and treats perception functions as black boxes by using a framework-independent format like ONNX.
Deploying data science code to vehicles requires bridging Python and C++, ensuring high code quality, and maintaining full traceability from requirements to artifacts.
msg
Ismaning, Germany
19:24Liang Yu
26:55Jan Zawadzki
38:12Daniel Graff & Andreas Wittmann
48:01Jan Zawadzki
30:49Marin Niehues
30:57Teresa Conceicao
27:23Nico Martin
1:00:28Tillman Radmer & Fabian Hüger & Nico Schmidt
Jobs that call for the skills explored in this talk.
Trinamics
Utrecht, Netherlands
autonomous-teaming
München, Germany
