Ground Truth and Labels
Every supervised result rests on labels assumed to be true; when that assumption is shaky, so is everything built on it.
The Assumed Bedrock
Ground truth is the reference against which a method is judged correct: the true class of an image, the true value of a measured quantity, the accepted answer to a benchmark. In supervised learning and in validation alike, ground truth is treated as bedrock. But it is often itself an estimate, a human judgment, or a proxy, and its imperfection sets a ceiling on everything above it.
Where Labels Come From
- Human annotation, which carries disagreement and error.
- Instruments, which have their own calibration and noise.
- Consensus of experts, which can be systematically biased.
- Proxies that stand in for the quantity actually wanted.
When Ground Truth Is Not True
Labels can be wrong, ambiguous, or inconsistent between annotators. A model trained to match noisy labels learns the noise; a benchmark scored against flawed answers rewards flawed behavior. A method cannot be more accurate than its reference labels, and reported accuracy above the label quality is measuring agreement with error, not correctness.
Proxies and Their Gaps
Often the true target cannot be measured, so a proxy is used. The result is then only as valid as the proxy's link to the real quantity. A model that predicts the proxy perfectly may still fail on the thing that actually matters if the two diverge, a gap that must be stated, not assumed away.
In Physics Validation
In science the analog of ground truth is experimental measurement, which carries its own uncertainty on both sides of any comparison. Honest validation reports error bars on the measurement as well as the simulation, so agreement is judged against real, not assumed-perfect, references. Kronos validation comparisons keep uncertainty on the experimental side visible for exactly this reason.