AI/ML development processes in the automotive environment
In machine learning projects, the software footprint is often modest: the model itself typically consists of fairly simple code, similar to what you'd find in relevant online forums. Things tend to get more involved with pre- and post-processing, which requires real programming skills. The ML model itself ultimately remains a black box that is trained and tested with data of varying quality.
Depending on how deep you want to go, the processes from Data Management SPICE can be used instead of process SUP.11. Higher data quality delivers considerable added value here.
The next topic is training and release for the respective use case. One important aspect that the SPICE model doesn't explicitly address: the data used for training must be strictly different from the data subsequently used to verify whether the model has genuinely learned all the edge cases.
All available, good-quality data is typically split into three groups: regular data for training the model, roughly 10 % "silver" data used during training to check whether the model has learned well, and a further 10 % or so of "gold" data for a final release check. If the check against the silver data still reveals weaknesses, targeted adjustments are made. If the check against the gold data reveals anything critical, the process starts completely over.
If you have projects where machine learning processes are the focus of the assessment, feel free to reach out. We're happy to support you with our expertise.
