Efficient Data Communication for Deep Learning Application via Latent Code Transmission in B5G Wireless Networks

In B5G networks, we introduce an innovative data transmission strategy, latent code transmission, tailored for applications that gather data from distributed devices and run deep learning algorithms on a centralized server. The essence of our approach is its adaptability to varying wireless channel conditions: it sends comprehensive original data during optimal channel conditions and, when conditions are suboptimal, processes the preliminary stages of the deep learning model at the device end, thereby shrinking the data size before transmission. The distinct advantage of our method is its ability to naturally reduce data size through the deep learning pipeline, obviating the need for an additional compression/decompression phase. This results in minimal extra computational demands compared to locally executing the deep learning model.