Interpreting Deep Neural Network-Based Receiver Under Varying Signal-To-Noise Ratios

TL;DR

This work addresses the challenge of interpreting a deep neural network–based wireless receiver (DeepRx) operating under varying SNR by introducing a post-hoc explainer that predicts channel parameters from internal activations. The method yields global and local explanations, identifying which units carry information about parameters like SNR and revealing how informativeness varies across layers and channels. Empirical results show middle layers are typically more informative for SNR, while certain channels may be pruned to compress the model without sacrificing performance; the approach also uncovers substantial instance-level variability and outliers. The study demonstrates robustness in high-dimensional settings and highlights generalizability to other architectures and channel parameters, with future work extending to real data and incorporating model fragility into the interpretability framework.

Abstract

We propose a novel method for interpreting neural networks, focusing on convolutional neural network-based receiver model. The method identifies which unit or units of the model contain most (or least) information about the channel parameter(s) of the interest, providing insights at both global and local levels -- with global explanations aggregating local ones. Experiments on link-level simulations demonstrate the method's effectiveness in identifying units that contribute most (and least) to signal-to-noise ratio processing. Although we focus on a radio receiver model, the method generalizes to other neural network architectures and applications, offering robust estimation even in high-dimensional settings.

Figures (8)

• Figure 1: Proposed method revealing how specific parts of the DeepRx model contain information about Signal-to-Noise Ratio processing.
• Figure 2: Interpreting the internal processing of the deep neural network-based receiver model (DeepRx) under varying Signal-to-Noise Ratios.
• Figure 3: Interpretations of selected ResNet layers and channels for the the deep neural network-based receiver model as performer model. Subplots (i) and (iii) show the means of the global interpretations across layers and channel-specific insights for layer B1-PRE, respectively, based on ten different data folds with standard deviations indicated by error bars. Subplot (ii) show local, instance-specific interpretations for channels 20 and 57 in layer B1-PRE. Activations in pre-activation ResNet blocks are denoted as follows: B1-PRE refers to activations after the first ReLU in the first ResNet block, while B1-POST refers to activations after the second ReLU in the same block. This naming convention continues similarly for subsequent blocks.
• Figure 4: Interpretations of selected ResNet layers and channels for the the deep neural network-based receiver model as performer model. Subplots (i) and (ii) show the means of the global interpretations across layers and channel-specific insights for layer B1-PRE, respectively, based on ten different random seeds with standard deviations indicated by error bars. The naming convention for layers follows that of Figure \ref{['fig:experimental_results']}.
• Figure 5: Cumulative contributions of test data instances to the global-level interpretations for the deep neural network-based receiver model as the performer model. Subplot (i) shows the cumulative contributions across layers, while subplot (ii) presents the cumulative contributions across channels for layer B1-PRE. The naming convention for layers follows that of Figure \ref{['fig:experimental_results']}. Contributions are displayed for the largest value, the ten largest values, the one hundred largest values, the one thousand largest values, and the remaining data instances. The contribution of each data instance is calculated as its individual value divided by the total sum of all data instances, reflecting its relative importance to the global-level interpretations.