Targeted Fine-Tuning of DNN-Based Receivers via Influence Functions

TL;DR

This work introduces influence functions to neural wireless receivers and applies them to DeepRx to identify training samples that most affect bit predictions. By selecting poorly performing targets and retraining on their most beneficial supporting samples, the method achieves BER improvements toward genie-aided benchmarks, outperforming random tuning in single-target scenarios. The approach leverages cross-loss alignment, relative influence, and Newfluence with scalable Arnoldi-based IHVPs, and demonstrates that single-target adaptation is feasible while multi-target adaptation remains challenging. Overall, influence-guided fine-tuning provides a data-efficient, interpretable pathway for adaptive, near-optimal neural receivers in dynamic wireless environments.

Abstract

We present the first use of influence functions for deep learning-based wireless receivers. Applied to DeepRx, a fully convolutional receiver, influence analysis reveals which training samples drive bit predictions, enabling targeted fine-tuning of poorly performing cases. We show that loss-relative influence with capacity-like binary cross-entropy loss and first-order updates on beneficial samples most consistently improves bit error rate toward genie-aided performance, outperforming random fine-tuning in single-target scenarios. Multi-target adaptation proved less effective, underscoring open challenges. Beyond experiments, we connect influence to self-influence corrections and propose a second-order, influence-aligned update strategy. Our results establish influence functions as both an interpretability tool and a basis for efficient receiver adaptation.

Figures (10)

• Figure 1: Illustration of the proposed approach: (top) neural receiver processes received signals to estimate transmitted bits with baseline performance; (bottom) targeted fine-tuning via influence functions leverages the most beneficial training instances, improving receiver performance and narrowing the gap to the genie-aided benchmark.
• Figure 2: Relative BER gaps $\Delta_{\text{BER}}$ as defined in Eq. \ref{['eq:relative_ber_gap']} between DeepRx and a genie-aided LMMSE (full CSI). Targets are the top-5 largest-gap instances in the regime where $\text{BER}_{\text{Genie}} \geq 10^{-3}$.
• Figure 3: BER over fine-tuning steps for first- and second-order methods with random and influence-guided selection: (a) most beneficial and (b) most harmful instances (BCE loss, $\ell$-relative, 10 seeds). Solid lines show mean; shaded areas indicate $\pm 1$ standard deviation.
• Figure 4: Fine-tuning results with first-order method and influence-guided selection of the most beneficial instances. Uncoded BER vs. SNR (55 bins) is shown for (a) single, 3 steps, 200k; (b) single, 3 steps, 800k; (c) multiple, 15 steps, 200k; and (d) multiple, 15 steps, 800k. Each plot compares baseline LMMSE, before fine-tuning, random fine-tuning, influence-guided fine-tuning, and genie-aided LMMSE. Tables report mean absolute BER and relative BER gap reduction for targeted and other validation instances (BCE, $\ell$-relative; calculated over 10 seeds).
• Figure 5: Relative reduction in the truncation error upper bound per added term ($k$), computed from schioppa2022scalingif with $\|\theta\|=1$. The dashed line marks $k=40$, the number of eigenvalues selected.