Linear Progressive Coding for Semantic Communication using Deep Neural Networks

We propose a novel linear progressive coding framework for obtaining hierarchical compressed representations (measurements) of data so that we can perform hierarchical-grain-level machine learning tasks timely and accurately using these representations. We first encode data into optimized low-rate coarse linear representations or measurements, which can be quickly communicated to the receiver and used for timely and accurate coarse-level classifications. We then design an additional set of optimized linear measurements or representations of the data so that the receiver can perform accurate finer-level classifications using these newly communicated representations together with the previously received coarse representations. Our proposed method can be considered as optimized hierarchical compressed learning or progressive semantic communications optimized for hierarchical-grain-level machine learning tasks, using low-cost linear measurements. Our experimental results on the MNIST and CIFAR-10 datasets show the linear progressive measurements enable timely performing coarse-level machine learning tasks with a small number of initial measurements, while for finer-level tasks, achieving overall accuracy and efficiency comparable to non-progressive methods.