Applications and Techniques for Fast Machine Learning in Science

Abstract

In this community review report, we discuss applications and techniques for fast machine learning (ML) in science-the concept of integrating powerful ML methods into the real-time experimental data processing loop to accelerate scientific discovery. The material for the report builds on two workshops held by the Fast ML for Science community and covers three main areas: applications for fast ML across a number of scientific domains; techniques for training and implementing performant and resource-efficient ML algorithms; and computing architectures, platforms, and technologies for deploying these algorithms. We also present overlapping challenges across the multiple scientific domains where common solutions can be found. This community report is intended to give plenty of examples and inspiration for scientific discovery through integrated and accelerated ML solutions. This is followed by a high-level overview and organization of technical advances, including an abundance of pointers to source material, which can enable these breakthroughs.

Keywords: big data; codesign; coprocessors; fast machine learning; heterogeneous computing; machine learning for science; particle physics.

Figure 1

The concept behind this review paper is to find the confluence of domain-specific challenges, machine learning, and experiment and computer system architectures to accelerate science discovery.

Figure 2

High-level overview of the stages in a GNN-based tracking pipeline. Only a subset of the typical edge weights are shown for illustration purposes. (A) Graph construction, (B) edge classification, and (C) track construction.

Figure 3

Simulated type Ia supernova light-curve and classification. Top: calibrated flux evolution in different DES band-passes as a function of normalized time (the first photometric measurement is set to time equals zero). Bottom: Baseline RNN classification probability evolution with respect of time, no host-galaxy redshift information was provided. At each photometric measurement, classification probability is obtained. The maximum light of the simulated supernova is shown in a gray dashed line and the simulated redshift of the supernovae is shown on the top z = 0.466. We highlight that redshift is not used for this classification but can improve results. Our baseline RNN classifies this light-curve as type Ia SN with great accuracy before maximum light, it only requires a handful of photometric epochs. (Möller and de Boissiére, 2019).

Figure 4

A 6GeV/c electron event in the ProtoDUNE detector. The x-axis shows the wire number. The y-axis shows the time tick in the unit of 0.5μs. The color scale represents the charge deposition.

Figure 5

Experimental 4D-STEM measurement of a dichalcogenide 2D material. Atomic map is inferred from the data, each diffraction pattern represents an average of 7 × 7 experimental images, green STEM probes are labeled for regions of the sample with one layer, vacuum, and two layers (Ophus, 2019).

Figure 6

The illustration of hardware-aware quantization and pruning. A given NN model can be compressed by using low precision quantization instead of single precision. The extreme case is to use 0-bit quantization which is equivalent to removing/pruning the corresponding neurons. The goal of compression is to find the best bit-precision setting for quantization/pruning to reduce model footprint/latency on a target hardware with minimal generalization loss.

Figure 7

Taxonomy of compute architectures, differentiating CPUs, GPUs and DPUs.

Figure 8

DPU architectures: Matrix of Processing Engines (MPE) on the left, and spatial architecture on the right.

Figure 9

FINN compiler flow.

Figure 10

Analog vector-by-matrix multiplication (VMM) in a crossbar circuit with adjustable crosspoint devices. For clarity, the output signal is shown for just one column of the array, while sense amplifier circuitry is not shown. Note that other VMM designs, e.g., utilizing duration of applied voltage pulses, rather than their amplitudes, for encoding inputs/outputs, are now being actively explored see, e.g., their brief review in Bavandpour et al. (2018).