Speaker
Description
Current deep learning based models at the LHC produce deterministic point estimates without any accompanying measure of epistemic uncertainty. Without this information, the system cannot determine when its predictions may be unreliable, particularly in rare or weakly sampled regions of feature space. This work investigates a Bayesian Neural Network architecture targeting the Level-1 Trigger, using inference-time Monte Carlo Dropout to enable real-time uncertainty quantification alongside standard classification. The resulting predictive variance provides an online indicator of model reliability, improving score calibration and recovering up to 3.1 percentage points of classification accuracy under degraded inputs. To assess the computational feasibility of real-time inference, we present an FPGA implementation and show that the Bayesian components add a modest latency overhead while keeping total inference time within the sub-microsecond L1 budget on both a dense and a graph backbone. Together, these results suggest that Bayesian inference is a viable route to uncertainty-aware jet tagging within Level-1 trigger constraints.