Evaluation of autonomous systems under data distribution shifts

TL;DR

This paper investigates how data distribution shifts affect autonomous systems, proposing metrics to define safe operational limits and highlighting the degradation of network accuracy beyond certain shift thresholds.

Contribution

It introduces distance metrics to quantify distribution shifts and establishes empirical thresholds for safe autonomous system operation.

Findings

Network accuracy degrades with increasing data shift

Distance metrics can predict safe operation thresholds

Beyond thresholds, system performance becomes unreliable

Abstract

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