Leveraging Multi-Task Learning to Model Task Relationships in Composite VLC Waveform Demodulation

Multi-task learning (MTL) has recently been applied to demodulation in visible light communication (VLC) systems. This work introduces a modulation-relationship-driven framework that extends prior research by treating each modulation stream within a composite waveform as a distinct task and by quantifying their empirical relationships. Using a cross-stitch (CS) architecture, the proposed framework jointly classifies beacon phase shift keying (BePSK) and beacon position modulation (BePM), quantifying inter-modulation coupling through a learnable <i>α</i>-matrix. The normalized coupling score of 0.71 provides an interpretable measure of task dependency that transforms feature sharing into a design axis. These results enable modulation-aware physical layer (PHY) waveform design suitable for resource-constrained VLC receivers and contribute to the development of explainable, NextGen VLC systems.