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 in Massive MIMO systems. While conventional methods rely on Compressed Sensing or Codebook-based techniques, Deep Learning has recently emerged as a powerful paradigm for this challenge. End-to-end (E2E) models, such as the framework proposed by Carpi et al. [ICC 2023], have demonstrated superior performance by jointly learning the Pilot Signals, the User’s Encoder, and the Base Station’s Decoder. Adopting a ``task-oriented approach", their objective maximizes the downlink sum-rate while minimizing a differentiable overhead penalty (controlled by a trade-off parameter $\lambda$). Crucially, their results demonstrate that this strategy achieves near-optimal performance with significantly compressed feedback, identifying an efficient operating regime well below the requirements of traditional reconstruction-based methods, which necessitate high-fidelity channel estimates. However, despite these performance gains, these 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 pipeline from Carpi et al. [ICC 2023], using this validated model as the faithful baseline for our analysis. We apply a novel methodological framework, combining Explainable AI techniques, such as Quantitative Density-Based Clustering, Dimensionality Reduction, and Contingency Analysis, to examine the system's internal strategy.Specifically, our analysis investigates the three fundamental learned stages of the pipeline: the Downlink Pilot Signals, the User's latent vector, a compressed representation extracted by the Encoder from the noisy, channel-distorted signals, and the final precoder vector, generated by the Base Station's network to execute the downlink transmission. We demonstrate that the network's internal strategy is not fixed but emerges as a direct function of the rate-overhead trade-off, effectively switching between two distinct operational modes:a Compression-Oriented Regime, where priority is given to compression efficiency (minimizing overhead) over rate maximization; a Performance-Oriented Regime, where priority is given to the maximization of the sum achievable rate.In the high-rate regime, the model aims for a faithful reconstruction of the channel to maximize the sum-rate. Ideally suited for channel estimation, it learns to generate orthogonal pilots at the sensing layer while focusing its energy on the specific channel sector. Here, the latent representations act as continuous, high-fidelity maps of the channel information, a characteristic mirrored in the continuous distribution of the precoders. Both these two continuous representations prove not to be clusterable.Conversely, in the balanced regime, the model prioritizes meaning over form. Forced to prioritize efficiency, it abandons orthogonality to evolve towards non-orthogonal, task-oriented pilots: the model learns to retain only task-useful information while discarding the rest. In this state, the Base Station outputs a finite set of clustered precoding vectors. Our clustering analysis quantifies this structure, identifying 11 stable prototypes for the user's latent space and 12-18 prototypes for the precoder. 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 deterministic mapping between the input latent and the output precoder). However, for ambiguous, high-interference scenarios, it dynamically switches to a context-aware policy (a ``one-to-many" scenario).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 ; secondly, it is entirely absent in the original, continuous, and 'non-clusterable' input channel. This demonstrates that the end-to-end system has learned to ``tessellate'' the continuous physical space: it partitions the non-clusterable channel manifold into discrete decision regions based on a rule of semantic proximity, mapping physically similar channels to the same prototype. Ultimately, this work demonstrates that E2E models are not inscrutable black boxes: it is possible to develop interpretations of the underlying learned functions. By unveiling these emergent behaviors, this thesis provides a new XAI-driven pathway for understanding, trusting, and validating learned communication systems.

Author dc:creator

• Roberta Bucchignani (24399740)

• Engineering
• Electronics and Electrical

Statement dc:rights

• In Copyright

DOI dc:identifier

https://doi.org/10.25417/uic.32994224.v1

OAI identifier oai:identifier

oai:figshare.com:article/32994224