Deep Feature Compression Method for Collaborative Intelligent Applications

Collaborative Intelligence is a new paradigm for realizing distributed deployment of Deep Neural Network DNNs, which is widely applicable to visual scene sensing applications for IoT systems. In this paradigm, the DNN model is divided into two parts, the edge submodel and the cloud submodel, which are deployed on the mobile edge device and the cloud server respectively, and the two collaborate to accomplish the inference task. In the collaborative inference process, the deep features generated by the edge sub-model need to be compressed and transmitted to the cloud sub-model through the wireless channel to reduce the transmission delay and transmission energy consumption. In order to improve the compression rate of deep features, a deep feature compression method (LCFC) based on lightweight convolutional neural network is proposed. A simple and efficient deep feature compressor is designed in this method, which utilizes a lightweight convolutional neural network and a uniform quantizer for channel dimension compression, spatial dimension compression, and quantization of deep features, which can significantly improve the compression rate of deep features and reduce the number of bits required for transmitting deep features. The proposed method is fully experimented on the CIFAR-100 image classification task, and the experimental results show that it improves the deep feature compression rate by up to 31.5% compared to the baseline method, while ensuring that the degradation of inference accuracy is less than 1%.