Correct-and-Generate SemCom: A closed-loop, confidence-calibrated decoder for training-free semantic communications

Du et al. (2023) show that training-free generative decoders can reconstruct content from minimal prompts, but instability in GAI generations undermines accuracy-sensitive use cases. Borrowing the idea of self-correcting functions from Gillman et al. (ICML 2024), this project proposes a closed-loop SemCom decoder that couples a generative model (e.g., diffusion or VAE-GAN) with a small “consistency critic” trained to map generated samples toward regions of higher posterior plausibility given the transmitted semantics. The system:

• Uses multi-modal prompts (as in Du et al.) but adaptively tunes the diffusion step count or decoder guidance strength based on conformal uncertainty estimates.
• Employs a corrector function (cf. Gillman et al.’s stabilizing “self-consuming” loops) that nudges outputs toward the true data distribution conditioned on the received semantic tokens, while an abstain-and-retransmit policy triggers only when calibrated uncertainty exceeds a threshold.
• Jointly optimizes over channel coding and friendly jamming parameters (per Du et al.), but with an additional objective for calibration sharpness and error-recovery latency.
  Novelty is in the closed-loop coupling of training-free generative decoding with model-based correction and real-time uncertainty calibration for on-device reliability—turning diverse GAI outputs from a source of instability into a controllable feedback signal. If successful, this could make SemCom practical in safety-critical settings (e.g., face transmission) without costly joint training.

References:

1. Generative Al-aided Joint Training-free Secure Semantic Communications via Multi-modal Prompts. Hongyang Du, Guangyuan Liu, D. Niyato, Jiayi Zhang, Jiawen Kang, Zehui Xiong, Bo Ai, Dong In Kim (2023). IEEE International Conference on Acoustics, Speech, and Signal Processing.
2. Self-Correcting Self-Consuming Loops for Generative Model Training. Nate Gillman, Michael Freeman, Daksh Aggarwal, Chia-Hong Hsu, Calvin Luo, Yonglong Tian, Chen Sun (2024). International Conference on Machine Learning.