New Machine Learning Architectures For Fast Data Selection

粒子物理实验可能会对他们的物理覆盖范围产生巨大的限制，因为产生的数据量非常大，因此它们可以存储的数据量是如此之少。此外，大型强子对撞机等设施产生的大量数据本质上不像稀有过程那样具有科学意义。为了解决这些挑战，开发了“触发”系统来快速扫描数据并识别潜在的感兴趣的签名。这些系统必须为感兴趣的过程保持高效率，同时也要保持在实验的技术范围内，例如延迟-算法做出决定所需的时间。在这项工作中使用了几种技术，如处于不同选择阶段的现场可编程门阵列、中央处理器或图形处理器，每种解决方案都更适合不同的任务。然而，每种解决方案都有自己的局限性。新的Xilinx Versal平台旨在将CPU、GPU和FPGA的优势结合到一个灵活的并行处理平台中。这种架构特别适合于基于机器学习(ML)的算法，我们建议利用我们在这一领域的专业知识来评估这些新平台，并在为未来的粒子物理实验开发新的ML架构方面领先一步

Particle Physics experiments can have huge limitations to their Physics reach due to the sheer amount of data that is produced, and consequently how little they can store. In addition huge amounts of the data produced at a facility such as the LHC are intrinsically not as scientifically interesting as rare processes. In order to solve these challenges "Trigger" systems were developed to quickly scan data and identify potentially interesting signatures. These systems have to maintain a high efficiency for the processes of interest while also staying in the technical confines of the experiment such as latency - the time it takes for the algorithm to make its decision. Several technologies have been used in this endeavour such as FPGAs, CPUs, or GPUs at various stages of selection, with each of the solutions better suited to different tasks. However each of the solutions have their own limitations. The new Xilinx Versal platform aims to combine the benefits of CPUs, GPUs and FPGAs into a flexible platform for parallel processing. This architecture is particularly well suited to machine learning (ML) based algorithms and we propose to exploit our expertise in this area to evaluate these new platforms and gain a head-start in developing new ML architectures for future particle physics experiments