Evaluation of Novel Fast Machine Learning Algorithms for Knowledge-Distillation-Based Anomaly Detection at CMS

TL;DR

The paper addresses the challenge of real-time, flexible anomaly detection in the CMS Level-1 Trigger by distilling a high-capacity CICADA autoencoder into small, fast student models. It introduces and compares three quantization strategies—HGQ, LGN, and a QKeras baseline—using quantization-aware training and outlier exposure to preserve anomaly-detection performance while drastically reducing resource use, enabling sub-200 ns FPGA inference. Open CMS data is used to demonstrate that HGQ and LGN match or surpass the baseline in ROC AUC while requiring far fewer hardware resources, facilitating finer calorimeter input granularity. The work has practical impact for deploying advanced ML in latency-constrained environments and offers broader relevance to fast ML on resource-limited hardware in current and future high-energy physics facilities.

Abstract

The CICADA (Calorimeter Image Convolutional Anomaly Detection Algorithm) project aims to detect anomalous physics signatures without bias from theoretical models in proton-proton collisions at the Compact Muon Solenoid (CMS) experiment at the Large Hadron Collider. CICADA identifies anomalies in low-level calorimeter trigger data using a convolutional autoencoder, whose behavior is transferred to compact student models via knowledge distillation. Careful model design and quantization ensure sub-200 ns inference times on FPGAs. We investigate novel student model architectures that employ differentiable relaxations to enable extremely fast inference at the cost of slower training -- a welcome tradeoff in the knowledge distillation context. Evaluated on CMS open data and under emulated FPGA conditions, these models achieve comparable anomaly detection performance to classically quantized baselines with significantly reduced resource usage. The savings in resource usage enable the possibility to look at a richer input granularity.

Figures (5)

• Figure 1: Data rate vs. latency requirements across various scientific and industrial applications.
• Figure 2: Schematic depiction of the topology of CICADA's inputs. The energy deposits on the unrolled cylindrical calorimeter are image-like data.
• Figure 3: ROC curves for the two simulated signal samples. The teacher, the baseline student and three selected novel student models are shown. On the x-axis, the trigger rate is shown which is proportional to the false positive rate.
• Figure 4: Performance results for the different CICADA student models given as a function of resource utilization calculated in EPOPs. Any given model lies along a vertical line. On the left, the EMD between the targets from the teacher and the students' prediction is shown. On the right, the ROC-AUC for two simulated signals is shown. The teacher achieved ROC-AUCs of 0.81 and 0.85 on the $t\bar{t}\rightarrow l+\nu+2q$ and $t\bar{t}\rightarrow4q$ signals, respectively.
• Figure 5: CICADA teacher model architecture.