Latent Space Alignment in AI-Native Communications: Theory & Applications

This thesis studies communication between independently trained artificial intelligence models from a latent-space perspective. We move beyond bit-centric abstractions and view AI-native communication systems as exchanging structured representations, where failures arise from geometric incompatibility between transmitter and receiver latent spaces. Accordingly, semantic noise is modeled as latent-space mismatch, that is, structural misalignment between heterogeneous representation spaces that prevents direct semantic interpretability. The thesis develops a unified framework for Semantic Channel Equalization (SCEq) in AI-native communications. Building on the concept of Relative Representations (RRs), we propose their use as a semantic equalization mechanism enabling zero-shot alignment between independently trained encoders, without joint retraining or model sharing. We then introduce a frame-based reformulation leading to the Parseval Frame Equalizer (PFE), yielding numerically stable reconstruction operators, controlled semantic compression, and robustness to perturbations via well-conditioned linear analysis–synthesis mappings. Moving beyond the zero-shot, semantic-only setting, we next consider the coupling of latentspace alignment with wireless transmission, formulating a joint physical–semantic equalization problem over Multiple-Input and Multiple-Output (MIMO) channels. Semantic precoding and decoding are optimized to jointly compensate channel impairments and latent-space misalignment. We study both linear and neural formulations, leading to tractable alternating optimization in the linear case and higher-capacity nonlinear alignment in the neural case. The approach is further integrated into Deep Joint Source–Channel Coding (DeepJSCC) architectures, where heterogeneous encoder–decoder pairs require additional alignment stages to preserve semantic consistency. Finally, we address dynamic, resource-constrained semantic communication scenarios via a stochastic optimization framework that jointly adapts communication, computation, and learning parameters, while preserving semantic alignment under latency and accuracy constraints. Overall, the thesis provides a mathematically grounded, system-level treatment of latent space alignment in AI-native communications, establishing semantic channel equalization as a central component for interoperability across heterogeneous intelligent agents.