Deep Learning for online tagging of proton-proton collisions at the High-Luminosity LHC

The Large Hadron Collider is a key facility in high-energy physics and one of the most prominent producers of big data. However, storing all this data is unfeasible; thus, a significant portion is filtered out using a trigger system. The upcoming High Luminosity upgrade is expected to increase data generation, demanding more efficient data processing strategies. Deep Learning is emerging as a promising approach for enhancing the trigger process. Deep Neural Networks can identify patterns in large datasets efficiently, making them suitable for early-stage data selection. In particular, Deep Neural Networks can be implemented on FPGAs used in trigger boards, offering the necessary speed and flexibility for real-time analysis. However, implementing those Deep Learning algorithms on FPGAs is challenging due to resource limitations. An efficient way of adapting Deep Neural Networks to these constraints is network pruning, which involves removing non-critical elements, thereby reducing network complexity without significantly affecting performance. This process typically involves training the network, removing the less critical elements, and retraining the pruned model to recover performance. This cycle of pruning and retraining is repeated to incrementally simplify the network, in a sub-optimal time-consuming process. Our research explores a novel pruning approach that focuses on removing unnecessary nodes in Deep Neural Networks without the need for retraining cycles. This method involves a “shadow network” that operates in parallel to the original network during training, identifying and deactivating superfluous nodes. This process leads to a pruned network optimized for efficient FPGA implementation, potentially advancing trigger solutions and enabling online tagging of collision outcomes at the High Luminosity phase of the Large Hadron Collider.