Logic Gate Neural Networks: Ultra-Fast ML for Triggers and Beyond
by
R61 CR03
RAL
Abstract:
What if neural networks were made of logic gates instead of neurons? Logic gate networks (LGNs) offer exactly this — replacing traditional nodes with learnable logic gates that yield implicitly pruned, discretized networks at inference time. The result: ultra-fast inference that's a natural fit for FPGAs and trigger systems. Better still, because the networks are pure logic, they open the door to mathematically provable explainability, which is a rarity in ML. This talk explores applying LGNs to anomaly detection at the CMS Level-1 Trigger, demonstrating competitive physics performance at a fraction of the inference cost. But the potential goes further: LGNs could enable trigger-level ML that wasn't previously feasible, with applications also extending to space systems and other latency-critical environments.