Speaker
Description
Future high-energy physics (HEP) experiments operate under extreme real-time constraints, where online filtering and trigger decisions increasingly define the ultimate physics reach. Although machine learning is now widely used in online systems, current deployments are almost exclusively limited to inference with offline-trained models. In this contribution, we investigate a complementary and largely unexplored paradigm: real-time learning directly in hardware, upstream of the trigger.
We present a fully digital hardware architecture for neural network training, explicitly designed for continuous operation in high-throughput online environments. Unlike conventional approaches, the architecture targets learning rather than inference, enabling autonomous adaptation to incoming data streams without reliance on pre-trained weights or strong a priori assumptions. This capability is particularly relevant for HEP, where unbiased data-driven learning before triggering offers a path toward improved sensitivity to rare, evolving, or unforeseen event signatures.
The complete training system is described, including forward and backward propagation, cost evaluation, parameter updates, and control logic, forming a self-contained learning primitive suitable for integration into online and front-end electronics. A proof-of-principle FPGA implementation validates the design. Crucially, the study provides realistic timing and resource estimates based on full implementation results, establishing concrete bounds on achievable network size, data absorption rate, and scalability on current-generation devices.
Beyond a single demonstrator, an applicability landscape is derived that maps online learning capacity to hardware constraints. This framework enables quantitative assessment of future real-time deployments and motivates a new class of adaptive online systems. Such learning-capable hardware blocks may become an essential component of next-generation trigger and online computing architectures for the HL-LHC and future experiments.