Threshold Shift
A defect type in which a systematic offset exists between the decision boundary learned by a model and a specified threshold. Behavior appears correct on the surface, but the model persistently misjudges near the boundary.
Timeline (last 90 days)
The study planted three types of defects in the training data: confounding factors that hide each rule, shifting the spinal threshold by 3 Gy, and implanting hidden triggers.
The 16 correctly identified networks included all 15 networks with confounding factors or threshold shifts.
Every confounding factor and threshold shift that was found was visible on a single neuron, and choosing one simple statistic for each defect type could also find the same defects.
All Facts (3)
The study planted three types of defects in the training data: confounding factors that hide each rule, shifting the spinal threshold by 3 Gy, and implanting hidden triggers.
50%UnverifiedThe 16 correctly identified networks included all 15 networks with confounding factors or threshold shifts.
50%UnverifiedEvery confounding factor and threshold shift that was found was visible on a single neuron, and choosing one simple statistic for each defect type could also find the same defects.
50%