Interpreting Neural Receivers: Mechanistic and Data Attribution Approaches

1 Introduction

Machine learning is expected to play an increasingly central role in future wireless communication systems. In particular, deep neural network-based radio receivers have been proposed as alternatives or complements to conventional signalprocessing pipelines in the physical layer of next-generation networks (e.g., 6G [Farhadi et al., 2025]). These models learn complex nonlinear signal transformations directly from data to recover transmitted information from noisy wireless signals, and have demonstrated strong performance in challenging propagation environments [Honkala et al., 2021] and alternative receiver paradigms such as pilotless communication [Korpi et al., 2023].

Despite these advantages, their adoption raises challenges related to interpretability. Deep neural networks often operate as black-box models whose internal decision processes are difficult to analyze. While interpretable machine learning has advanced significantly [Kim et al., 2023], most methods have been developed for domains such as vision and language. Physical-layer machine learning differs substantially: inputs are non-semantic signals, outputs are high-dimensional, and deployment constraints are strict. These challenges motivate interpretability approaches tailored to neural receivers. This research investigates how representation-level analysis and data attribution can reveal model behavior, relate learned representations to physical channel properties, and support engineering tasks such as monitoring and adaptation. While motivated by wireless communications, the work contributes to broader AI questions related to mechanistic interpretability, data attribution, and robustness under distribution shift.

2 Research Problem

Interpretability is inherently context-dependent [Kim et al., 2023], and methods effective in one domain may fail in another. Neural receivers present several distinctive challenges. First, inputs consist of complex-valued signals without direct semantic meaning, limiting the usefulness of standard input attribution methods. Second, models produce large vectors of outputs rather than single class predictions, making classwise explanations impractical. Third, strict deployment constraints limit methods that modify model architectures or add computational overhead.

This research addresses these challenges by developing post-hoc, model-specific interpretability methods that analyze internal representations and training data influence without modifying the architecture. The work considers neural receiver architectures derived from DeepRx [Honkala et al., 2021], a setting in which interpretability remains largely unexplored despite growing industrial interest in AI-native communication systems and trustworthy AI. DeepRx is a convolutional neural receiver for orthogonal frequency division multiplexing systems used in modern communication standards such as 5G NR and emerging 6G. The model takes received noisy signal samples as input and outputs estimates of the transmitted bits along with their associated confidence.

3 Research Contributions

The research studies interpretability of neural receivers through four complementary directions: mechanistic interpretability, training data attribution, out-of-distribution detection, and targeted model adaptation.

Mechanistic Interpretability. This work investigates how internal representations of neural receivers encode physically meaningful channel parameters. A post-hoc explainer model is trained in a supervised manner to predict channel characteristics from internal representations of a trained receiver [Tuononen et al., 2025b]. The predictive performance of the explainer is used as a proxy for the information content of individual layers and convolutional channels. Experiments show that signal-to-noise-ratio (SNR) information is not uniformly distributed across the network; instead, informativeness varies significantly across convolutional channels. These results demonstrate that neural receivers learn structured internal representations aligned with physical channel properties, and suggest opportunities for architecture analysis, pruning, and model simplification.

A related study analyzes the geometry of internal representations under varying channel conditions [Tuononen et al., 2026a]. The results show that activations lie on a smooth lowdimensional manifold strongly aligned with SNR, providing insight into how neural receivers organize signal information and highlighting the role of continuous structure in representation learning, as opposed to discrete semantic classes.

Out-of-Distribution Detection. Building on these mechanistic insights, this work investigates how internal representations can be used to detect distribution shifts. A layerwise out-of-distribution detection framework is developed using channelwise aggregation of internal representations [Tuononen et al., 2026a]. In-distribution behavior is modeled using statistics computed from training data, and deviations from this structure are used to detect anomalous inputs. Experimental results show that representationbased detectors can reliably identify certain types of distribution shifts, such as extreme multipath conditions, while other shifts such as high mobility remain more challenging. The analysis further shows that earlier network layers often provide stronger signals for distribution shift detection than deeper layers. These findings demonstrate how representation-level interpretability can inform the design of monitoring mechanisms for neural receivers.

Data Attribution and Model Adaptation. A complementary perspective is obtained through training data attribution methods that analyze how individual training samples influence model behavior [Tuononen et al., 2026b]. In this work, influence functions [Koh et al., 2017] are used to operationalize this perspective by estimating how removing specific training samples affects model predictions. Experiments show that influence-guided selection of fine-tuning data can accelerate adaptation to new channel conditions compared to random selection. Performance is evaluated as a reduction in the bit-error rate (BER) gap between the neural receiver and a genie-aided linear minimum mean square error (LMMSE) benchmark. These results demonstrate the practical value of data attribution for efficient and targeted model adaptation.

Methodological Contributions. This work addresses numerical instability in k-nearest neighbor-based normalized mutual information estimation in high-dimensional settings through a logarithmic reformulation [Tuononen et al., 2025a]. The approach improves numerical stability and enables more reliable information-theoretic analysis of internal representations, supporting mechanistic interpretability by quantifying their dependence on channel parameters.

4 EvaluationMethodology

The proposed methods are evaluated using DeepRx-based neural receivers trained on simulated wireless channel data and tested under both in-distribution and out-of-distribution conditions. Performance is measured using communication metrics such as BER, alongside interpretability metrics derived from representation and attribution analysis. Ongoing work extends evaluation to over-the-air wireless measurement data, focusing on deployment-oriented scenarios.

5 Impact and FutureWork

The increasing integration of machine learning into communication infrastructure raises important questions of reliability, transparency, and accountability in safety- and performance-critical systems. Interpretability methods address these challenges by enabling understanding of model behavior, detection and diagnosis of failures, and adaptation under changing conditions. By improving transparency and supporting informed decision-making, such methods contribute to the responsible deployment of neural receivers and to a more trustworthy and equitable use of this technology.

Future work will investigate the generality of learned representation structures in neural receivers, particularly their role in detecting and characterizing distribution shifts, and develop improved data attribution methods for model adaptation. This includes advancing information-theoretic analysis of neural representations and exploring underlying computational mechanisms. At a more practical level, future work will examine how interpretability can be integrated into systemlevel validation, monitoring, and adaptation workflows.

More broadly, interpretability may serve not only as a diagnostic tool but also as a mechanism for discovery, providing insights that can inform the design of future communication systems and inspire new signal processing approaches.

Acknowledgments

The author gratefully acknowledges the support of the Nokia Foundation and Nokia Corporation.