Data-Driven Tuning of Deep Unfolding-Aided Blind Signal Separation for Short Data Length

This paper proposes deep unfolding (DU)-aided blind signal separation (BSS) for short data block communications of multi-user multi-input multi-output (MU-MIMO) transmission. Independent component analysis (ICA) is a BSS technique that does not require a pilot signal for channel estimation. However, when applying ICA to short data length, inaccurate non-Gaussianity evaluations deteriorate the performance of BSS. To compensate for this drawback, we design a novel DU-aided BSS algorithm as trainable ICA (TICA) that incorporates a data-driven design based on deep learning and learning optimization of internal trainable parameters. Considering that the weight matrix obtained as the outcome of ICA does not converge to an appropriate value for short data length, we introduce two trainable parameters: one for correcting the non-Gaussian evaluation and the other for damping to suppress oscillations. Numerical results show that the proposed algorithm substantially improves the bit error rate (BER) performance with trainable parameter optimization using DU techniques.