LoRa modulation for split learning

In this paper we introduce a task-oriented communication design for split learning (SL) over a communication channel. Our approach involves the Expressive Neural Network (ENN), a novel neural network featuring adaptive activation functions (AAF) based on the Discrete Cosine Transform (DCT). This architecture does not only provide better learning capabilities, but also facilitates data transmission using the Long Range (LoRa) modulation. The frequency nature of LoRa is adequate for the communication side of the problem, while allowing to construct the AAFs at the receiver. Additionally, we propose orthogonal chirp division multiplexing (OCDM) for multiple access and a modified modulation aimed at preserving communication bandwidth. Our experimental results demonstrate the effectiveness of this scheme, achieving high accuracy in challenging scenarios, including low signal to noise Ratio (SNR) and absence of channel state information (CSI) for both additive white Gaussian noise (AWGN) and Rayleigh fading channels.
This work is part of the project IRENE (PID2020-115323RB-C31), funded by MCIN/AEI/10.13039/501100011033 and supported by the Catalan government through the project SGR-Cat 2021-01207.
Peer Reviewed
Postprint (author's final draft)