Signal Compression for Efficient Partitioning of Deep Neural Networks

Fifth generation (5G) mobile networks are adopting several techniques to provide higher data rates while meeting strict latency requirements. Advanced compression techniques and distributed data processing are among them supporting cost efficient network deployments. Also, the usage of machine learning techniques to optimize telecommunication systems is gaining momentum, specially after promising results with deep learning models. Following this trend, there are investigations towards splitting the processing of such models between different network nodes, making these applications more suitable and adaptable to scenarios with low processing capacity nodes. Therefore, in this work we investigate and propose different compression techniques for efficient partitioning of deep neural networks. We combine several splits with quantization and Huffman coding compression algorithms, providing insights on the configurations with the best performance. We compress the output score of the model to reduce the overhead of transmitting such scores through the network and evaluate how the accuracy is affected by this compression, and which compression technique provides the best performance.