[ABT-1] Integrating Communication, Sensing, and Learning Toward the Digital-Twin Era: Semantic Communications, Representation Learning, and Optimal System Design

3:10 PM - 3:50 PM JST(6:10 AM - 6:50 AM UTC)

[ABT-1-03] Latent Representation Learning toward Wireless Environment Digital Twins: CSI Sensing via Distillation and Contrastive Learning

〇Takayuki Nishio¹ (1. Science Tokyo)

Keywords:

CSI Sensing,WiFi Sensing,Representation Learning,Latent Distillation,Contrastive Learning

Machine-learning-based CSI sensing for integrated sensing and communication (ISAC) suffers from strong environment dependence and the high cost of label collection. This paper focuses on unsupervised representation learning that acquires general-purpose CSI representations without labels, and reviews its foundations—latent distillation and contrastive learning. We further organize recent trends in CSI representation learning into reconstruction-based, latent-prediction-based, and contrastive-based approaches, and present the authors' case studies that enable sensing under few-sample, missing-input, and environment-varying conditions: enhancement of robustness to missing features, cross-modal learning using image latent representations as teachers, and high-resolution image generation via latent representation estimation (LatentCSI).