Uncertainty

Uncertainty, in the context of a machine learning model or estimate, is a measure of how confident that prediction actually is, as distinct from the prediction's raw value. A model can output a prediction without any indication of how reliable it is; explicitly quantifying uncertainty allows a system (or a person) to know when to trust a prediction and when it needs closer review.¹

Why flagged uncertainty is more useful than silent, unreliable confidence

A prediction with high uncertainty flagged as such is more useful than a confident-looking but silently unreliable one, which is why quantifying uncertainty, not just producing a prediction, matters for a system meant to support human review of edge cases.

Where uncertainty applies

Not on the drawing; uncertainty is a property of a model's output when applied to a specific part or situation, not something the part's requirements specify.

Common mistakes

Treating every model prediction as equally reliable regardless of context; a model is often less certain for unusual geometry or situations far from its training data, and that difference matters for how much the prediction should be trusted.

Is uncertainty the same as confidence?

They're closely related; uncertainty and confidence are often two ways of describing the same underlying reliability estimate, sometimes as complementary measures.

Why would a manufacturing AI system need to quantify uncertainty?

To flag predictions on unusual parts or situations for human review, rather than presenting every output with equal, unwarranted confidence.

Does high uncertainty mean a prediction is wrong?

Not necessarily; it means the model has less confidence in that specific prediction, which is useful information regardless of whether the prediction turns out correct.