Learning-Based Coded Computation
Learning-Based Coded Computation
复制标题
DOI：
10.1109/jsait.2020.2983165
复制
发表时间：
2020-03
期刊：
影响因子：
--
通讯作者：
J. Kosaian;K. V. Rashmi;S. Venkataraman
中科院分区：
文献类型：
--
作者：
J. Kosaian;K. V. Rashmi;S. Venkataraman
最近的进展表明，编码计算的潜力，赋予弹性的减速和故障发生在分布式计算系统。然而，现有的编码计算方法要么不能支持非线性计算，要么只能支持有限的非线性计算子集，同时需要高的资源开销。在这项工作中，我们提出了一个基于学习的编码计算框架，以克服对一般非线性函数进行编码计算的挑战。我们表明，在编码计算框架内仔细使用机器学习可以扩展编码计算的范围，从而为更一般的非线性计算提供弹性。我们展示了基于学习的编码计算对神经网络推理的适用性，这是生产服务中的主要工作量。我们的评估结果表明，基于学习的编码计算能够准确重建广泛部署的神经网络的各种推理任务（如图像分类，语音识别和对象定位）的不可用结果。我们在一个开源预测服务系统上实现了我们提出的方法，并显示了它在减轻神经网络推理中出现的减速方面的承诺。这些结果表明，潜在的学习为基础的方法打开新的大门，使用编码计算更广泛的，非线性计算。
Recent advances have shown the potential for coded computation to impart resilience against slowdowns and failures that occur in distributed computing systems. However, existing coded computation approaches are either unable to support non-linear computations, or can only support a limited subset of non-linear computations while requiring high resource overhead. In this work, we propose a learning-based coded computation framework to overcome the challenges of performing coded computation for general non-linear functions. We show that careful use of machine learning within the coded computation framework can extend the reach of coded computation to imparting resilience to more general non-linear computations. We showcase the applicability of learning-based coded computation to neural network inference, a major workload in production services. Our evaluation results show that learning-based coded computation enables accurate reconstruction of unavailable results from widely deployed neural networks for a variety of inference tasks such as image classification, speech recognition, and object localization. We implement our proposed approach atop an open-source prediction serving system and show its promise in alleviating slowdowns that occur in neural network inference. These results indicate the potential for learning-based approaches to open new doors for the use of coded computation for broader, non-linear computations.