Machine Learning qualification handbook Flashcards
Purpose of the Machine Learning Qualification Handbook
- provide guidelines on
- how to create reliable AI functions
- and perform validation & verification (V&V)
- considering the specifics of AI development practices in the space domain.
Which organizations are part of the working group for the Machine Learning Qualification Handbook?
Airbus (Convenor)
ESA
Mathworks
Spacebel
Ariane Group
CNES
DLR
What areas are covered under the Guidelines Scope of the Machine Learning Qualification Handbook?
Bottom-Up-Approach:
- Data Qualification & ML Model Development Process
- Model Testing
- System Testing & Qualification
What are the key aspects of data qualification process?
Data qualification
- Define high-level data quality properties
- Understand data lifecycle
- Assess risks associated with different data types: real, simulated & synthetic, and augmented
- Address application and data specific needs for different learning types (supervised, unsupervised, reinforcement)
- Operational scenarios and data relevance
What are the methods and focuses of model testing in ML qualification?
- Testing operational scenarios and operational design domain
- Emphasizing ML interpretability and explainability
- Testing methods:
- Specific example testing
-
NN coverage testing:
- ensure that during testing various parts of NN are activated
- discover edge cases or potential errors that may not surface under normal usage
-
Out of domain testing:
- test ML model on data that differs from data it was trained on, but which might be encountered in a real world scenario
- Crucial for assessing model robustness as it helps to identify how the model behaves under unexpected or novel conditions
- particularly important in safety-critical applications like those in the space domain
-
noise testing:
- introducing various types of noise or perturbations to the input data to check how resilient the model is against such disturbances.
- E.g.: adding random noise to images, sound clips, or other input data types to simulate real-world imperfections
- adversarial testing: evaluating a model’s security and robustness by trying to fool it using deliberately crafted inputs,
- Single Event Upset testing:
What does system testing and qualification involve in ML qualification?
- Analyze ML software interaction with the system and compliance with system requirements
- Define levels of autonomy (human assisting function, human machine collaboration, full autonomy)
- Focus on Category B, C, D
- FMEA/FMECA for identifying potential failures and mitigation strategies
What is FMEA and what does it involve?
Failure Modes and Effects Analysis is a systematic method for identifying potential failure modes in a product or process, assessing the risks associated with those failures, and determining actions to mitigate them. It involves:
- Identifying how each component can fail. (What Can Fail How?)
- Analyzing the effects of these failures. (Causes what?)
- Rating the severity, occurrence, and detection of failures. (How bad, how likely, how hard to find?)
- Calculating a Risk Priority Number (RPN) to prioritize mitigation efforts based on risk. (Which priority?)
What is FMECA and what does it entail?
Failure Modes, Effects, and Criticality Analysis extends FMEA by adding a criticality analysis, quantifying the severity and likelihood of failure effects to prioritize them based on risk. It includes:
- Conducting an FMEA to identify failure modes and effects.
- Assigning criticality values based on severity and occurrence.
- Ranking failure modes by criticality to focus mitigation efforts.
- Developing detailed actions for the highest risks to enhance system reliability and safety.
What is a Single Event Upset?
Why is it crucial the effects of SEUs in the context of Machine Learning & AI systems used in space?
Single Event Upset:
- change of state caused by ions or electromagnetic radiation striking a sensitive node in a microelectronic device
- radiation can cause the device to behave unpredictably.
SEU testing in the context of machine learning and AI systems:
- assess how well a system can tolerate and recover from these random, unpredictable changes in its hardware components
- helps ensure that the ML model can continue to operate correctly despite potential disruptions caused by radiation effects (a common challenge for spacecraft electronics)
What are the key aspects of ML model development process?
- Model quality characteristics
- Model selection process
- Controlling for common pitfalls in ML model training