AdaStop addresses a practical gap in deep neural network testing: deciding when to stop labeling test inputs. The framework models each labeling action as incurring cost c, while discovering a fault provides value v. It estimates the marginal fault discovery rate during testing and stops when that rate falls below the decision threshold τ = c/v. Across multiple datasets, architectures, and test-selection strategies, the authors report that AdaStop discovers 65–84% of faults using only 9–31% of the labeling budget.
No heat snapshots are available in the last 24 hours.