Data-driven Optimization for Zero-delay Lossy Source Coding with Side Information

This paper proposes a data-driven architecture for zero-delay lossy source coding with side information (i.e., Wyner-Ziv coding) for sources with memory. The overall architecture involves designing suitable filters at the encoder and the decoder and performing fixed-rate scalar quantization followed by one-dimensional binning of quantization indices. Unlike previous work, which uses an exhaustive search to optimize the system parameters, this paper proposes a lower-complexity data-driven method that does not require a priori knowledge of source and side information statistics. The main ingredients of the proposed approach include modeling the quantization process by an additive quantization noise process, modeling the modulo operation by a continuous approximation, and approximating the decoding process by a softmin function, which makes the system amenable to training using stochastic gradient descent. Experimental results on Gauss-Markov sources with different memory orders demonstrate that our proposed system can match the performance of systems optimized using an exhaustive search.