Average output across sampled subnetworks.
Predictive paths across familiar and unfamiliar inputs
training observationsstochastic passespredictive meanquery
ACT · uncertainty is below thresholdThe agent can use this prediction under the selected uncertainty policy.
Model disagreement at the query position.
Approximate mean plus or minus 1.96 standard deviations.
Policy result under the selected abstention threshold.
Dropout becomes an ensemble
At inference, keep dropout active and repeat the forward pass. Every mask selects a slightly different subnetwork. The collection approximates a distribution over predictions rather than a single point estimate.
Disagreement signals model uncertainty
When sampled subnetworks agree, epistemic uncertainty is low. When they diverge, the model may be outside well-supported regions. This is distinct from irreducible observation noise.
Abstention turns uncertainty into behavior
An agent needs a policy, not just a score. Calibrate thresholds on held-out data, measure risk and coverage, and route high-uncertainty actions to a safer tool, a human, or more evidence.
Measure the whole decision system.
Separate uncertainty types
MC dropout primarily targets epistemic uncertainty. If labels or sensors are noisy, model aleatoric uncertainty explicitly and evaluate both contributions.
Calibrate the threshold
Predictive variance is not automatically a probability of failure. Choose abstention rules against task loss, accepted-case accuracy, coverage, and important cohorts.
Budget the passes
More passes stabilize estimates but add latency. Benchmark convergence, batch the stochastic passes when possible, and stop increasing samples when decisions stop changing.