Martin takes the audience back to basics when it comes down to discussing Machine Learning (ML) in Automotive Safety. He explains that ML is a subset of Artificial Intelligence (AI), and its application requires a specific strategy.
About Martin Stock, SGS-TÜV Saar GmbH: Martin is Senior Safety Expert at SGS-TÜV, where he works as a trainer of both ISO 26262 and ISO 21448. He also has a vast experience in SOTIF and is a member of the IEEE P2851 and ISO/TS 5083 working groups.
Machine Learning (ML) and Artificial Intelligence (AI). Two closely-related, yet different terms that seem to be used interchangeably in the world of technology and innovation. However, an analysis of the basics of these concepts is needed to guarantee successful safety projects. So, what are those basic definitions that have been altered by blurry misconceptions, and what is the strategy for industry applications?
“Machine Learning is a subset of Artificial Intelligence,” says Martin Stock, automotive safety expert at SGS-TÜV. On one hand, he explains, AI is the ability of computer systems –respectively E/E systems– to solve problems using cognitive skills that simulate those of humans. On the other, ML is the study of computer algorithms that improve automatically through experience and by the use of data.
In a broader sense, AI is based on cognitive skills to detect and differentiate situations even when information is missing or ambiguous. More specifically, in ML, systems learn and improve from experience (E) with respect to some class of tasks (T) and performance (P).
Use of ML in Automotive Safety and Strategy
According to ISO/DIS 21448 (SOTIF), ML is used for object detection and classification in automated driving functions. Martin underlines that “Algorithms, based on types of ML, are mainly used when a full specification of the ‘problem at hand’ (e.g., fully specified and characterized types of traffic lights or pedestrians) is not available or possible.”
Martin further asserts, “Successful ML algorithms shall be trained on valuable data which can only be derived appropriate when the Operational Design Domain (ODD) and the Object and Event Detection and Response (OEDR) are specified and understood sufficiently.”
Strategy for ML Training Process
When setting up an ML training process strategy, the initial step is to define the functions implemented by one or more Dynamic Driving Tasks (DDTs). Once this is done, safety experts need to specify ODD & OEDR by defining the specific conditions under which the Dynamic Driving Tasks (DDT) –especially the Object and Event Detection and Response (OEDR)– is designed to function.
A key step in Martin’s formula is to derive training examples based on the ODDs, DDTs, and OEDRs, to finally being able to generate training data based on such examples via the various learning techniques: Decision and regression trees, classification rules, neural networks, clustering techniques, genetic algorithms, etc.
Once the data is generated, it needs to undergo a training data process to further validate it through experts who can evaluate its plausibility, or through online validation by testing the learned data in a field application. In this whole process, Martin underlines: “The development of training data is susceptible regarding systematic faults. Therefore, the specification of a ‘training process’ or ‘training strategy’ by experts is recommended.”
This content has been presented during The ISO 26262 Digital Conference.