Know Your Limits: Machine Learning with Rejection for Vehicle Engineering

Abstract

New vehicle designs need to be tested in representative driving scenarios to evaluate their durability. Because these tests are costly, only a limited number of them can be performed. These have traditionally been selected using rules of thumb, which are not always applicable to modern vehicles. Hence, there is a need to ensure that vehicle tests are aligned with their real-world usage. One possibility for obtaining a broad real-world usage overview is to exploit the data collected by sensors embedded in production vehicles. But these do not produce the detailed data needed to derive the metrics computed using expensive sensors during testing. Therefore it is necessary to correlate the low-end sensor measurements available in production vehicles with the relevant metrics acquired using high-end sensors during testing. Machine learning is a promising avenue for doing this. The key challenge is that vehicles will be used "in the wild" in many scenarios that were not encountered in the controlled testing environment, and it is unlikely that learned models will perform reliably in these previously unseen environments. We overcome this challenge by allowing learned models to abstain from making a prediction when unexpected vehicle usage is identified. We propose a general framework that combines standard machine learning with novelty detection to identify previously unseen situations. We illustrate our framework's potential on data we collected from a large-scale roadroughness analysis use case. Empirically, our approach can identify novel road types in the wild and by doing so it yields better performance.

A3SOM, abstained explainable semi-supervised neural network based on self-organizing map

“…One potential improvement to our model would be to include abstention during the training phase. In the survey [83], three architectures for abstained classification are described: separated, dependent, or integrated rejection. We perform dependent rejection: abstention is applied to the output of the model, after the training step.…”

Detecting Evasion Attacks in Deployed Tree Ensembles

“…For such examples, the model's predictions have an elevated risk of being incorrect, and hence may not be trustworthy. An example can be rejected due to ambiguity (i.e., how well the decision boundary is defined in a region) or novelty (i.e., how anomalous an example is with respect to the observed training data) [22]. The OC-score metric goes beyond measuring ambiguity in an ensemble (i.e., the model's confidence in a prediction).…”

Can We Trust Fair-AI?

“…An underdeveloped research line consists of rejecting the output of an AI system in favor of escalating the decision to a human agent who could possibly take into account additional (qualitative) information. This is considered in the area of classification with a reject option (or selective classification) (Hendrickx et al 2021). There is a trade-off here between the performance of an AI system on the accepted region, which should be maximized, and the probability of rejecting, which should be minimized (as human agents' effort is limited).…”