On Improving Validity of Deep Neural Networks in Safety Critical Applications

Licentiatavhandling, 2020

Context: Deep learning has proven to be a valuable component in object detection and classification, as the technique has shown an increased performance throughput compared to traditional software algorithms. Deep learning refers to the process, in which an optimisation process learns an algorithm through a set of labeled data, where the researcher defines an architecture rather than the algorithm itself. As the resulting model contains abstract features retrieved through the optimisation process, new unsolved challenges emerge that need to be resolved before deploying these models in safety critical applications.

Aim: The aim of this Licentiate thesis has been to study what extensions are necessary to verify deep neural networks. Furthermore, the thesis studies one challenge in detail: how out-of-distribution samples can be detected and excluded.

Method:
A comparative framework has been constructed to evaluate performance of out-of-distribution detection methods on common ground. To achieve this, the top performing candidates from recent publications were used as a reference snowballing baseline, from which a set of candidates were studied. From the study, common features were studied and included in the comparative framework. Furthermore, the thesis conducted semi-structured interviews to understand the challenges of deploying deep neural networks in industrial safety critical applications.

Results: The thesis found that the main issue with deployment is traceability and quality quantification, in the form that deep learning lacks proper descriptions of how to design test cases, training datasets and robustness of the model itself. While deep learning performance is commendable, error tracing is challenging as the abstract features in the do not have any direct connection to the training samples. In addition, the training phase lacks proper measures to quantify diversity within the dataset, especially for the vastly different scenarios that exist in the real world.

One safety method studied in this thesis is to utilize an out-of-distribution detector as a safety measure. The benefit of this measure is that it can both identify and mitigate potential hazards. From our literature review it became apparent that each detector was compared with different techniques, hence a framework was constructed that allowed for extensive and fair comparison. In addition, when utilizing the framework, robustness issues of the detector were found, where performance could drastically change depending on small variations in the deep neural network.

Future work: Future works recommend testing the outlier detectors on real world scenarios, and show how the detector can be part of a safety strategy argumentation.