Interpretation of learned solutions for Precoding-oriented Massive MIMO CSI Feedback Design

The development of effective channel feedback strategies is crucial for enabling the widespread adoption of Frequency Division Duplexing (FDD) in Massive MIMO systems. Conventional methods for multi-user FDD Massive MIMO with limited feedback fall into two families: Compressed Sensing and Codebook-based techniques. Recently, Deep Learning has emerged as a powerful paradigm for the challenging downlink FDD massive MIMO CSI feedback problem, outperforming conventional methods. End-to-end (E2E) models, such as that of Carpi et al. [ICC 2023], have demonstrated superior rate-overhead performance by jointly learning the pilot signals, the user's encoder and the base station's precoder. They adopt a "task-oriented approach": their objective is to maximize the downlink sum-rate while simultaneously minimizing a differentiable overhead penalty, with the balance between the two controlled by a trade-off parameter λ. However, these high-performance models operate as opaque "black boxes," leaving their learned internal strategy unknown. This thesis provides an original interpretability study to "open" this black box. As the authors' code was not public, we first successfully replicated the E2E learning pipeline from Carpi et al. [ICC 2023], using this validated model as the faithful baseline for our interpretability analysis. We apply a novel methodological framework, combining Explainable AI techniques such as quantitative density-based clustering, dimensionality reduction for visualization, and contingency matrices, to the internal learned representations. Specifically, we analyze two key stages of the pipeline: the user's latent vector, which is the learned representation of the channel generated by the user's encoder, and the resulting precoder vector, generated and learned by the Base Station's network. We demonstrate that the network's internal strategy is not fixed but emerges as a direct function of the rate-overhead trade-off. Our initial analysis quantifies this dependency: in the high-compression (low λ) regime, the precoder collapses to a single, fixed vector while the latent vector shows minimal diversity. As λ increases, the number of unique vectors for both the latents and the precoders explodes exponentially, confirming the model learns richer, continuous representations when the constraint is relaxed. This observation prompted our main finding: when the system is forced to prioritize high compression, this limited diversity is not a simple collapse, but the formation of a finite, 'emergent codebook.' Our clustering analysis quantifies this structure, identifying 11 stable prototypes for the user's latent space and 12-18 prototypes for the precoder. We provide twofold proof that this is a learned strategy, not an inherent data property. Firstly, the structure completely dissolves as the compression constraint is relaxed (at high λ); secondly, it is entirely absent in the original, continuous, and "non-clusterable" input channel. By decoding the mapping from the latent clusters to the precoder clusters, we uncover the Base Station's learned policy: for simple channel states, it adopts an efficient lookup table (a 1-to-1 mapping). However, for ambiguous, high-interference scenarios, it dynamically switches to a context-aware policy (a 1-to-many mapping). This work demonstrates that E2E models can develop interpretable, finite, and semantically meaningful strategies, providing a new XAI-driven pathway for understanding and validating learned communication systems.