ETON Talk: Deep-Free Physics-Informed Foundation Modeling in Wireless Communications (slides)
This talk revisits classical generative modeling approaches, with an emphasis on mixture models developed prior to the rise of deep generative learning, and highlights their continued relevance for the physical layer of wireless communications. Owing to their capacity for analytical inference, mixture models form a natural connection between machine learning and classical signal processing, where covariance information is central to many physical-layer tasks.
By incorporating the underlying physics of wireless channels, these models enable lightweight, interpretable, and easily trainable solutions with strong generalization properties. The framework naturally extends to multimodal settings through multi-view mixture models, providing a unified and efficient approach for handling diverse data types. Overall, the talk bridges classical methods with modern generative modeling and demonstrates that effective generative modeling need not rely solely on deep learning.
