Imagine a conveyor belt carrying 30 million parcels every second, and a controller who must instantly decide what is inside: a valuable shipment or junk. At the Large Hadron Collider, that conveyor belt is the trigger of the LHCb experiment, picking out rare particle decays among billions of collisions. Previously, the controller was a set of simple rules: if a signal exceeded a threshold, the particle looked like a muon or an electron. But the rules let through a lot of junk, and precious signals sometimes got lost.

Now neural networks are replacing the rules: trainable controllers that see not just one feature but the whole picture. To keep the network from being distracted by noise and mistaking random fluctuations for signals, its sensitivity was constrained: even if the input changes slightly, the output does not flinch. It is like a controller with a steady hand who does not startle at every rustle. The constraint is imposed through weight normalization: each weight in the first layer is individually bounded, and in subsequent layers the sum of weights entering a neuron is bounded. GroupSort activations do not increase sensitivity, so the network's behavior remains predictable.

Training was done on simulated events: the signal consisted of J/ψ and Υ(1S) decays into lepton pairs, while the background was protons for muons and pions with kaons for electrons. Input features were scaled to the range from 0 to 1. For muons, the network received five variables: the chi-square of matching the track to hits in the muon stations, the sum of squared residuals, and three differences in track slope between subdetectors. For electrons, there were six variables: two estimates of the energy-to-momentum ratio, the squared distance between the extrapolated track and the cluster center, and three shower shape moments in the calorimeter. The networks turned out to be tiny: about 500 parameters, which is roughly 2 kilobytes in single precision. For each track, they add only 24–28 bytes of data.

The result exceeded expectations. At the same signal efficiency, the neural networks suppress background better than the previous threshold algorithms in all studied kinematic regions. For muons, the gain is largest at moderate and high momenta, where most tracks pass after selection; at low momenta the difference is small. For electrons, the improvement is most noticeable at low momenta, where the calorimeter energy resolution is worse and the single E/p feature is weak. The network is almost independent of pseudorapidity and transverse momentum. Profiling a typical HLT1 sequence shows that the muon and electron identification kernels each consume 0.5–0.9% of the total GPU time, less than the threshold operators they replace.

Before this work, it was unclear whether neural-network lepton identification could fit into the tight GPU trigger budget without losing speed or reliability. Now it has been shown that it can. By constraining the network's sensitivity through weight normalization and GroupSort activations, and then training it on simulated signals and backgrounds, background rejection was improved in all kinematic regions while using less than 1% of GPU time. This approach can be applied to any streaming event selection system that needs fast and robust classification, from other detectors at colliders to industrial sensors. The controller on the conveyor belt has become smarter, but has not lost its steady hand.