ML safety requirements assurance

Objectives

1. Develop the machine learning safety requirements from the allocated system safety requirements.
2. Validate the machine learning safety requirements against the allocated safety requirements, the system and software architecture and operational environment.
3. Create an assurance argument, based on the evidence generated by meeting the first two objectives, that provides a clear justification for the ML safety requirements. This should explicitly explain the tradeoffs, assumptions and uncertainties concerning both the safety requirements and the process by which they are developed and validated.

Inputs to the stage

• [E] : Safety requirements allocated to the ML component
• [I] : ML safety requirements argument pattern

Outputs of the stage

• [H] : ML safety requirements
• [J] : ML safety requirements validation results
• [K] : ML safety requirements argument

Description of the stage

As shown in the AMLAS ML safety requirements assurance process diagram above, this stage consists of three activities that are performed to provide assurance in the ML safety requirements. The artefacts generated from this stage are used to instantiate the ML safety requirements assurance argument pattern as part of Activity 5. The scope of this stage is limited to the ML model (e.g. the mathematical representation of the neural network that produces the intended output).