Inverse-GMM: A Latency Distribution Shaping Method for Industrial Cooperative Deep Learning Systems
Inverse-GMM: A Latency Distribution Shaping Method for Industrial Cooperative Deep Learning Systems
复制标题
DOI：
10.1109/jsac.2022.3229448
复制
发表时间：
2023-03
影响因子：
16.4
通讯作者：
Fei Qin;Yucong Xiao;Xian Sun;X. Dai;Wuxiong Zhang;Fei Shen
中科院分区：
文献类型：
--
作者：
Fei Qin;Yucong Xiao;Xian Sun;X. Dai;Wuxiong Zhang;Fei Shen
前端部署的深度学习是下一代工业应用的一项很有前途的技术，它可以从高维传感器中提取必要的信息。然而，在资源受限的前端设备上的这些繁重的计算任务的一部分必须被卸载到边缘或云设备，这通过中间数据的交换形成了协作深度学习系统。协作深度学习系统的推理效率将与非平稳工业多径衰落信道引起的通信延迟高度相关。本文提出了一种控制通信延迟分布的新方法，该方法能够在恶劣的工业环境中支持高效的协作深度学习架构。该方法本质上是高斯混合模型（GMM）的逆过程，它调整延迟样本以逼近给定的任意形状函数。为了实现这一目标，本文提出了一种新的解析域EM算法，将任意分布形状分解为多个高斯核，并提出了一种优化的随机资源分配算法来近似每个高斯核。经典的莱斯信道模型和现场实测的工业衰落信道响应验证了该方法的性能。
The front deployed deep learning is a promising technology of the next generation industrial applications, which can extract essential information from high dimension sensors. However, part of these heavy computation tasks at resource constrained front devices have to be offloaded to the edge or cloud devices, which forms the cooperative deep learning system through the exchange of intermediate data. The inference efficiency of cooperative deep learning system will then be highly correlated with the communication latency caused by the non-stationary industrial multipath-rich fading channel. This paper proposes a novel method to control the distribution of communications latency, which is able to support efficient cooperative deep learning architecture in the harsh industrial environment. The proposed method is essentially an inverse process of Gaussian Mixture Model (GMM), which adjusts latency samples to approach the given arbitrary shape function. To achieve this objective, a new variation of Expectation-Maximization (EM) algorithm in analytical domain is derived to decompose arbitrary distribution shape with multiple Gaussian kernels and an optimized stochastic resource allocation algorithm is proposed to approximate each Gaussian kernels. The performance of proposed method is verified by both classical Rician channel model and field measured industrial fading channel responses.