Neural-Network-Aided Engineering: New Frontiers in Automation

由于近年来计算能力和数据的可用性趋同，深度神经网络（DNN）受到了极大的关注。深度神经网络擅长机器学习（ML）：深度神经网络中的许多层可以在许多尺度上学习模式。然而，更多的层也意味着：训练需要更多的数据；而且，推理需要更长的时间。计算技术的进步，从移动设备到仓库规模的集群，意味着深度神经网络不再不切实际。此外，从未有过如此多的数据。DNN的时代已经到来。*** ***我们之前开发了OPAL，它通过使用深度学习来自动设计和优化DNN，以确定DNN结构和行为与DNN性能之间的关系。在基准数据集上验证了OPAL优化精度和GPGPU推理能量。我们建议大幅扩展OPAL，以自动设计(1)机器学习性能和成本受限的物联网应用的深度神经网络，(2)可信赖的深度神经网络，可以容忍输入不规则和硬件故障，以及(3)在功率、散热和可靠性约束下优化的嵌入式多处理器片上系统（MPSoC）。*** ***首先，我们将在OPAL中开发一个通用框架，用于指定和实现新的目标（或成本）函数。在设计(a)实时约束或(b)没有高性能GPGPU的任何东西时（即大多数物联网系统），需要替代成本函数。我们将研究预测多个相关目标的效果；然后，我们将针对新的目标优化深度神经网络，包括加速器、实时、嵌入式和可配置系统。*** ***其次，我们将设计度量标准来衡量以下影响：(a)缺失或不确定的输入数据；(b)推理过程中的硬件故障；对DNN性能的影响这些问题出现在真实的数据和环境中，但没有被基准数据集捕获。我们将开发测量准确度变化的方法，补偿这些问题，并最终优化系统以容忍它们。*** ***第三，我们将使用OPAL设计MPSoC。给定目标平台，设计人员将应用程序任务分配给资源（映射），并随后对其执行进行排序（调度）。今天特别重要的是在热或系统寿命限制下进行设计；我们假设OPAL将有效地学习空间任务映射与由此产生的热效应和寿命效应之间的关系，从而更快地找到更好的解决方案。*** ***这项研究对HQP和加拿大的好处是显而易见的。参与HQP将获得DNN和支持基础设施的宝贵经验，为不断增长的机器学习/物联网领域做好准备。蒙特利尔和QC ML的努力也将受益：我们的工作将改进他们现有的解决方案，并通过减少ML和物联网开发所需的资源，简化其他产品和服务的引入。

Deep neural networks (DNN) are receiving tremendous attention because of the recent convergence in the availability of computing power and data. DNN excel at machine learning (ML): the many layers in DNN can learn patterns at many scales. However, more layers also implies: more data is required for training; and, inference takes longer. Advances in computing, from mobile devices to warehouse-scale clusters, means DNN are no longer impractical. Furthermore, there's never been more data. The era of DNN has arrived.*** ***We previously developed OPAL, which automatically designs and optimizes DNN by employing deep learning to determine the relationship between: DNN structure and behavior, and DNN performance. OPAL has been demonstrated optimizing accuracy and GPGPU inference energy on benchmark datasets. We propose to significantly extend OPAL to automatically design (1) DNN for ML performance- and cost-constrained applications in the IoT, (2) trustworthy DNN that can tolerate input irregularity and hardware failure, and (3) embedded multiprocessor systems-on-chips (MPSoC) optimized under power, thermal, and reliability constraints.*** ***First, we will develop a general framework in OPAL for specifying and implementing new objective (or, cost) functions. Alternative cost functions are needed when designing (a) for real-time constraints, or (b) anything without a high-performance GPGPU: i.e., most IoT systems. We will investigate the effect of predicting multiple, related objectives; we will then optimize DNN for new targets, including accelerators, real-time, embedded, and configurable systems.*** ***Second, we will devise metrics for measuring the effect that: (a) missing, or uncertain, input data; and, (b) hardware failure during inference; have on DNN performance. Such concerns arise in real data and environments but are not captured by benchmark data sets. We will develop methodology for measuring the resulting changes in accuracy, compensating for such issues, and ultimately optimizing systems to tolerate them.*** ***Third, we will use OPAL to design MPSoC. Given a target platform, designers assign application tasks to resources (mapping), and subsequently order their execution (scheduling). Of particular importance today is design under thermal or system lifetime constraints; we hypothesize that OPAL will efficiently learn the relationship between spatial task mapping and the resulting thermal and lifetime effects, finding better solutions faster.*** ***The benefits of this research for HQP and Canada are clear. Participating HQP will gain valuable experience with DNN and supporting infrastructure, positioning them well for the growing ML/IOT domain. Montreal and QC ML efforts will also benefit: our work will improve their existing solutions, and simplify the introduction of additional products and services by reducing the resources required for ML and IoT development.