面向协作智能应用的深度特征压缩方法

Deep feature compression method for collaborative intelligence applications

协作智能是一种实现深度神经网络(Deep Neural Network,DNN)分布式部署的新范式,广泛适用于物联网系统视觉场景感知应用。在该范式中,DNN模型被划分为边缘子模型和云端子模型两部分,分别部署在移动边缘设备和云服务器上,由两者协作完成推理任务。协作推理过程中需要对边缘子模型生成的深度特征进行压缩,通过无线信道传输给云端子模型,以降低传输延迟和传输能耗。为了提高深度特征的压缩率,提出了一种基于轻量级卷积神经网络的深度特征压缩方法(LCFC)。该方法设计了一种简单高效的深度特征压缩器,它利用轻量级卷积神经网络和均匀量化器对深度特征进行通道维度压缩、空间维度压缩和量化,能显著提升深度特征的压缩率,减少用于传输深度特征所需的比特数。所提出的方法在CIFAR-100图像分类任务上进行了充分实验,实验结果表明,在保证推理精度退化小于1%的情况下,与基线方法相比,该方法将深度特征压缩率最高提升了31.5%。

Collaborative Intelligence(CI) is a new paradigm of Deep Neural Network(DNN) distributed deployment,which is widely applied to IoT visual scene perception application.In this paradigm,the DNN model is partitioned into two parts:the edge sub-model and the cloud sub-model,which are deployed in the mobile edge device and the cloud server respectively.They cooperate to complete the inference task.In the process of cooperative inference,the deep features generated by the edge sub-model should be compressed before being transmitted through the wireless channel,so as to reduce the transmission delay and transmission energy consumption.In order to improve the compression rate of deep features,a deep feature compression method based on lightweight convolutional neural networks is proposed in this paper.In this method,a simple and efficient deep feature compressor is designed.It uses lightweight convolutional neural network and uniform quantizer to compress and quantify deep features respectively,which can significantly improve the compression rate of deep features and reduce the number of bits required for transmitting deep features.The proposed method has been fully tested on the CIFAR-100 image classification task,and the experimental results show that compared with the baseline method,the deep feature compression rate of our method can be improved by up to 31.5% when the inference accuracy degradation is less than 1%.