Encouraging Validatable Features in Machine Learning-Based Highly Automated Driving Functions

As more Highly Automated Driving (HAD) functions are implemented using Machine Learning (ML)-based methods, the challenge of validating them is undeniable. In a prior work, we proposed a validation method which analyzes the feature embeddings of Deep Neural Network (DNN) classifiers. Using this method, if different DNNs with similar classification performance are given, an engineer can inspect the feature embeddings and choose the DNN showing the most meaningful embeddings. This is a form of validation of the chosen architecture. In our prior work, the feature embeddings were passively observed with the goal of choosing the architecture with the most meaningful embeddings. In this work, we modify the DNN loss function in order to encourage more meaningful feature embeddings, aiming to actively strengthen the validation of a given DNN architecture. To this end, we make use of k-means friendly spaces, introduced in the context of autoencoders. We argue that these lead to desirable feature embeddings for validation. Furthermore, we introduce two classification rejection rules, which can be used to reject certain classifications. This increases the overall performance of the ML-based method. Ultimately, these rejection rules positively benefit from the k-means friendly space. We use a lane change prediction task as a safety-critical HAD function use-case throughout the paper. We show that the proposed methods can be used on a wide range of ML-based algorithms.