In this chapter you can find an overview of several data-driven techniques for feature learning of radio-frequency (RF) signals. From shallow dimensionality reduction to deep metric learning and self-supervised learning, we discuss different approaches to leveraging channel estimates for wireless localization.

We propose a self-supervised method that learns general-purpose channel features from unlabeled data without relying on contrastive CSI estimates. Furthermore, we investigate varying Transformer attention settings to leverage antenna and subcarrier diversity.

We show that the whole estimated channel can be fed into the transformer block as a set of subcarriers. Without vectorizing the input, using recurrence, standard convolution operators, or fusion, attention can serve as an adaptive filter for resilient CSI.

The fronthaul overhead issue in distributed RAN for 5G and 6G is addressed in a learning-based approach based on extreme value theory.

The work investigates and proposes MCD and Deep ensembles to account for the uncerstainty measure in wireless channel and network parameters.

Whether contrastive triplet-loss is better than classical cross-entropy based classifier for low-dimensional CSI representation learning.