Evaluation of autonomous systems under data distribution shifts
This addresses safety concerns for autonomous systems in real-world applications, but it is incremental as it builds on existing ideas about distribution shifts with a simple demonstration.
The paper tackles the problem of autonomous systems failing under data distribution shifts by proposing that safety requires halting operation beyond a threshold, demonstrating with a toy example that network accuracy degrades with shifts and using distance metrics to define safe limits.
We posit that data can only be safe to use up to a certain threshold of the data distribution shift, after which control must be relinquished by the autonomous system and operation halted or handed to a human operator. With the use of a computer vision toy example we demonstrate that network predictive accuracy is impacted by data distribution shifts and propose distance metrics between training and testing data to define safe operation limits within said shifts. We conclude that beyond an empirically obtained threshold of the data distribution shift, it is unreasonable to expect network predictive accuracy not to degrade