Machine learning or AI models. A system of quality metrics for evaluating different types of uncertainty, confidence, or suitability in the ML model being used. Exceeding one or more of these dimensions could be an indication of a model-breaking event.
Degree of Uncertainty (Certainty)
- Your uncertainty of things you have models for based on your observations and actions
- Subjective to model and task
Degree of Ignorance (Complete Knowledge)
- The incompleteness of your models for the task at hand, the number of surprises
- Subjective to surprises from model
Degree of Generality (Special-Case Knowledge)
- Specific vs General representation of concepts, their reusability, and legibility
- Subjective to entropy/compactness (Schmidthuber) of representation and effective task transferability
Degree of Vagueness (Concreteness)
- Fuzzy properties of model
- Degrees of truth
- Concreteness vs. vagueness
Degree of Obstinance (Uncommitted, Waffling)
- Sluggishness of model adaptation when environment changes
- Subjective to rate of model change relative to surprisal
Degree of Myopia (Open-Minded)
- Overcommitment to a single view or state of the environment
- Inability to hold conflicting or broader hypotheses at the same time that may be relevant later
- Subjective to multiple task satisfaction, task switching, robot kidnapping recovery, and resistance to catastrophic forgetting, no overfitting
Degree of Incompetence (Expertise)
- Inability to effect desired changes in the environment under perception
- (Polani)
- Subjective to percepts and effectors
- task-satisfaction success or failure