Neural joint source-channel coding via Bernoulli latent straight-through estimator

Under infinite block length, from Shannon's separation theorem, it is well-known that by independent design of source coding and channel coding, the optimal throughput — capacity can be reached with careful corresponding functionalities. However, when restricted to finite theorem does not necessarily hold, hence joint source-channel (JSCC) has been raised as an alternative strategy for achieving channel. JSCC formulates highly non-convex problem which cannot directly solved analytically, recently, deep learning proposed a key enabler JSCC. While most work on deep-learning-based focused transmission continuous signals through wireless channels, we focus in this paper case information-theoretic channel, namely binary symmetric give useful insights perspective Unlike recent deep-learning-based, or neural, discrete channels considered score function estimator train JSCC, paper, improve performance neural estimating gradient more precisely compared previous approach. Key idea consider soft codeword during training enable path-wise proven have lower variance than score-function estimator. Experimental results MNIST CIFAR-10 datasets show outperforms validating effectiveness computation technique.