This question is very important and more focused on reccomendations and lessons learnt. In fact, the “issue” of “legibility” of ML, in terms of being able to explain in human terms how the system behaves, is unsolved (Dosovitskiy et al., 2015) and (Zeiler et al., 2014). This makes difficult to predict how “classical” V&V techniques can be applied to ML systems. For this reason, reccomendation and guidelines can be absolutely helpful.
Main Question
Have safety recommendations been identified?
Sub-Questions
- Are any tools used to evaluate how well the model performs for each type of prediction?
- Are any safety-relevant “lessons learned” recognized and stored?
- Can they be shared (when necessary, so internal to the company, or other OEMs or Suppliers, Regulatory bodies, and so on)?
References
- Dosovitskiy, A. and Brox, T. (2016) ‘Inverting Visual Representations with Convolutional Networks’. [online] ArXiv. doi:https://doi.org/10.48550/arXiv.1506.02753.
- Zeiler, M.D. and Fergus, R. (2013) Visualizing and Understanding Convolutional Networks. [online] ArXiv. Available at: https://arxiv.org/abs/1311.2901.