Adaptive Unequal Error Protection for Semantic Split Learning over Wireless Channels

What the math gives to ML

The paper turns semantic relevance into a per-coordinate protection signal by estimating mutual information between a latent representation and the downstream target, then ranking latent coordinates by the gradient of that estimate. The transferable asset is not the wireless PHY implementation itself, but a task-aware mechanism for deciding which learned features should receive lower noise, higher precision, more bits, or greater redundancy. This can be transplanted into split inference, feature dropout, quantization, pruning, and latent-channel training by making corruption strength inversely proportional to estimated task relevance. The key experiment is to compare MI-guided corruption or resource allocation against uniform allocation at equal communication cost and equal latent dimensionality.