Autoencoder-Based Unequal Error Protection Codes

We present a novel autoencoder-based approach for designing codes that provide unequal error protection (UEP) capabilities. The proposed approach, based on generalization of an autoencoder loss function, provides versatile framework the design message-wise and bit-wise UEP codes. Using associated weight vector, generalized function can be used to trade off probabilities corresponding different importance classes explore region achievable probabilities. For UEP, we compare with union random coset are compared rateless spinal superposition Gaussian In all cases, show superior performance while providing simplicity flexibility in trading among classes.