Foundation models improve communication quality at ultra low data rates
Foundation Models for Generalizable Semantic and Goal-Oriented Communication
Machine LearningArtificial IntelligenceRobotics
Summary
Sending images using very limited data often causes important details to be lost, especially when the system only learns from a small set of examples. The authors propose a new method that uses large pretrained vision and language models to pick out key parts of an image to send, and then uses a special diffusion model to fill in the missing parts on the receiving end. This approach helps avoid overfitting to training data and keeps the meaning of images intact even when very few bits are sent. Their experiments show improved performance compared to other methods, including when tested on images not seen during training.
semantic communicationgoal-oriented communicationfoundation modelsdiffusion modelsimage compressionbit rateoverfittingsemantic fidelity
Authors
Boliang Liu, Wint Yi Poe, Riccardo Trivisonno, Giuseppe Caire
Abstract
Semantic and goal-oriented communication is increasingly studied for 6G, but generalization beyond seen data remains a key weakness under tight rate budgets. Many existing systems overfit their training data and degrade sharply at very low bit rates because they attempt to compress the entire signal. We introduce Foundation Model-Guided Semantic and Goal-Oriented Communication (FMSGOC), a framework that uses broad visual-linguistic Foundation Model priors to mitigate overfitting. It further improves rate efficiency by concentrating bits on sparse, goal-aligned anchors and relying on generative foundation-model priors to reconstruct the masked regions. By decoupling what to send from how to reconstruct, a vision-language foundation model selects and transmits a sparse set of semantic anchors, while a pretrained diffusion model, fine-tuned for masked completion, reconstructs the image at the receiver. In our experiments, FMSGOC reaches 0.039 bits per pixel (BPP), maintains high semantic fidelity (cosine similarity 0.87-0.90 on CIFAR-10), remains robust on previously unseen inputs (0.83-0.86 on ImageNet), and shows good perceptual similarity (0.1278/0.1558, CIFAR-10/ImageNet), outperforming strong end-to-end baselines at lower bit rates.