Track papers across labs, topics, and trends.
Search papers, labs, and topics across Lattice.
This paper introduces REACH, an interpretability framework that enhances the performance of deep learning channel estimators for vehicular communications by identifying key time-frequency features and compressing the architecture. By employing input-level and filter-level attributions, the authors demonstrate that relevant features can be isolated to reduce input dimensionality while maintaining performance, and reveal a universal internal representation that explains improved out-of-distribution generalization. The resulting architecture compression achieves significant reductions in parameters and FLOPs with minimal degradation in accuracy, particularly under OOD conditions.
Identifying just a handful of critical features can lead to substantial architecture compression without sacrificing performance in vehicular channel estimators.
Multi-channel mixed-SNR training improves out-of-distribution (OOD) generalisation of deep learning channel estimators for IEEE 802.11p vehicular communications, yet the internal mechanism responsible for this remains unexplained. This work presents REACH (Relevance-based Explanation and Architectural Compression for cHannel estimators), a gradient-based interpretability framework that operates at two levels. Input-level attribution identifies a subset of time-frequency features consistently relevant across all evaluated channel conditions, enabling input dimensionality reduction with minimal performance loss. Filter-level attribution reveals a near-universal internal representation, providing a representational account of the observed OOD generalisation. Guided by the resulting filter taxonomy, relevance-guided architecture compression substantially reduces both the number of parameters and the number of floating-point operations (FLOPs) with sub-1 dB normalised mean square error (NMSE) degradation, and OOD generalisation degrades more slowly than within-distribution accuracy under increasing compression.