The Deployment of Realtime ML in Changing Environments

The High-Luminosity LHC upgrade of the CMS experiment will utilise a large number of Machine Learning (ML) based algorithms in its hardware-based trigger. These ML algorithms will facilitate the selection of potentially interesting events for storage and offline analysis. Strict latency and resource requirements limit the size and complexity of these models due to their use in a high-speed trigger setting and deployment on FPGA hardware. It is envisaged that these ML models will be trained on large, carefully tuned, Monte Carlo datasets and subsequently deployed in a real-world detector environment. Not only is there a potentially large difference between the MC training data and real-world conditions but these detector conditions could change over time leading to a shift in model output which could degrade trigger performance. The studies presented explore different techniques to reduce the impact of this effect, using the CMS track finding and vertex trigger algorithms as a test case. The studies compare a baseline retraining and redeployment of the model and episodic training of a model as new data arrives in a continual learning context. The results show that a continually learning algorithm outperforms a simple retrained model when degradation in detector performance is applied to the training data and is a viable option for maintaining performance in an evolving environment such as the High-Luminosity LHC.