Integrated Sensing and Learning for Better Generalized Edge AI

In recent years, significant advancements in deep learning, wireless communication, and sensing have laid the foundation for integrated sensing and learning (ISAL), which involves machines actively and collaboratively collecting data from the environment to facilitate model training at the network edge, specifically for tactile intelligence service provisioning. Despite the progress in deep learning, a it faces a vital issue of overfitting, wherein models excel on training samples but struggle with unseen ones, particularly when resource constraints are in play. To address this issue, we draw inspiration from the classic stochastic gradient Langevin dynamics (SGLD) approach, where a right amount of noise is introduced to gradients to alleviate overfitting and enhance model generalizability. We propose an over-the-air federated stochastic gradient descent (Air-FedSGD) scheme for distributed model training. This scheme inherently introduces the required noisy gradient akin to SGLD, where the noise level is jointly determined by the devices' transmission power and sensing duration. Within this context, we formulate a joint sensing and communication (SC) resource allocation problem with the objective of minimizing the population loss of the learned model. Unlike the commonly used empirical loss, population loss measures model performance not only on the training set but on every set identically independent of the training set, thereby giving us a handle on the generalization ability of a model. The solution to this problem establishes an AI task-oriented joint sensing and communications design framework, which is elaborated considering a specific use case of human motion recognition. Extensive experimental results validate the superiority of the proposed design, affirming its effectiveness in addressing over-fitting challenges and enhancing generalization capabilities.