acceptodds
Under review as a conference paper at ICLR 2027

Frozen Networks, Better Decisions: SNR-Conditioned Readout for Modulation Recognition

Abstract

Automatic modulation classification identifies how information is encoded in a received wireless signal. Noise can hide the patterns that distinguish modulation formats, making accurate recognition difficult. We ask how much accuracy can be gained after a classifier has been trained. We study affine readouts that adjust its output probabilities while keeping the network fixed. The maps are fitted on validation data and can use signal-to-noise ratio (SNR) to adapt their decisions to signal quality. This conditional form requires SNR at prediction time. In a post-hoc study across five public benchmarks, we compare these readouts on matched hard-label models and models trained with SNR-dependent soft targets. On the two related RadioML2016 benchmarks, independently fitted bin maps improve mean accuracy by about ten percentage points without retraining the network. This gain occurs in a validation-defined transition between low and high baseline accuracy. Sharing parameters across bins adds little on average. Gains are smaller on other benchmarks, and some conditional maps perform worse than a global map that does not use SNR. Equal readout treatment also leaves some training advantages intact outside the transition band. Simulated SNR errors can erase the readout gains. These results identify substantial room to improve trained modulation classifiers, while showing why training improvements should be evaluated against fitted output rules with their information requirements made explicit.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
automatic modulation classification, SNR-conditioned readout, affine calibration, frozen networks, training evaluation, SNR robustness

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_NCDpOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
3
Read review
Paper summary

This paper seeks to boost the accuracy of models trained for automatic modulation classification of communication signals by utilizing well-known calibration techniques to tune the readout as a function of the signal-to-noise ratio (SNR). However, these techniques are also shown to be sensitive to errors in SNR estimates. The contribution is narrow and, in my opinion, not impactful.

Critical strengths

The paper shows that, for a fixed ML model, SNR-dependent affine readouts can boost classification performance. Essentially, this applies known calibration techniques to each SNR bin to improve performance.

Critical weaknesses

The paper makes a simple and rather obvious observation, applying calibration to the problem of automatic modulation classification (AMC), but specialized to SNR bins. This does not meet the bar for a non-trivial contribution to a core ML conference. A better venue might be a conference focused on wireless communication.

Even within with narrow domain of AMC, no effort seems to be put into the fundamental problem of how to be robust to SNR variations in terms of model training and calibration as a whole. SNR estimates are, in general, not available for the AMC problem, so the problem of estimating SNR (explicitly or implicitly) must be convincingly addressed in order to contemplate using SNR-sensitive readouts.

The paper is poorly written in terms of technical clarity. For example, a soft target training scheme called DPC is used, but the expansion for the acronym, and more importantly, the details of the training methodology are not provided.

Questions for authors

You have shown that errors in SNR estimates adversely impact the performance of SNR-sensitive readouts. How do you propose obtaining accurate enough SNR estimates in an AMC problem?

Reviewer_ZU88OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper investigates SNR-conditioned post-hoc readout calibration for automatic modulation recognition (AMC). The authors freeze pretrained models and fit affine transformations of their predicted log-probabilities, conditioned on SNR bins. The study compares global, independently fitted, and shared SNR-conditioned readouts, alongside several training and calibration variants, including a soft-target training procedure.

Critical strengths

Improving an existing AMC classifier without retraining its backbone is valuable in settings where retraining is expensive or impractical.

Critical weaknesses

1.The presentation and writing are poor. The paper jumps directly into the proposed method without first providing a precise problem statement or a clear motivation section.

2.The motivation of the work is not clear. The central method relies on privileged information at inference time. The SNR-conditioned readout uses the true benchmark SNR to select its calibration branch. In realistic blind AMC deployment, this quantity is generally unknown and must itself be estimated from the received signal. The paper does not evaluate an end-to-end pipeline using a real SNR estimator.

3.The proposed per-bin affine transformation of log-probabilities is closely related to existing multiclass calibration methods, particularly Dirichlet calibration.

4.This article seems to be more like an experimental report and does not meet the standards of ICLR scientific research papers.

Questions for authors

1.Can the authors evaluate the complete inference pipeline using an actual SNR estimator operating on received IQ samples, rather than ground-truth SNR or synthetically perturbed oracle values? How do the results change under biased, class-dependent, or distribution-shifted estimation errors?

2.What is the specific methodological contribution beyond applying established affine or Dirichlet-style calibration separately across SNR bins? Is the shared readout statistically and practically better than independently fitted per-bin readouts?

3.All of the aforementioned Weaknesses.

Reviewer_QcJwOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The study explores the extent to which classification performance can be further improved through post-training output calibration, while keeping the underlying network frozen. It compares global and SNR conditioned affine readouts, including independent, shared and smoothed variants, alongside Structured Matrix Scaling. Across five benchmarks and 57 benchmark backbone combinations, the largest transition band accuracy gains occur on the two RadioML2016 datasets. Furthermore, the research investigates which differences between hard-label training and Distribution-Preserving Calibration training persist after applying identical output calibration, and examines how simulated SNR estimation errors affect these performance gains.

Critical strengths
  1. The study identifies a useful evaluation baseline. Independently fitted SNR-bin readouts improve hard-model transition-band accuracy by 11.30 percentage points on RadioML2016.10B and 9.39 points on RadioML2016.10A without changing the backbone. The consistent improvements across the 26 corresponding cells show why fitted output rules deserve consideration when evaluating training improvements.

  2. The experimental comparisons and interpretations are generally careful. Both training arms receive matched readout procedures, and the paper distinguishes remaining training differences from readout interactions. It reports accuracy, NLL and Brier score, retains negative outcomes, and explicitly acknowledges the post-hoc study design, additional SNR information and limited benefit of parameter sharing.

Critical weaknesses
  1. Independent confirmation and the scope of the empirical finding remain limited. Nested validation and the close agreement between validation and test gains support the observed improvements within the studied benchmarks. However, Section 4 acknowledges that earlier test results informed the study and that validation data had already participated in backbone checkpoint selection. The follow-up therefore does not provide a fresh, previously unexamined confirmation of the finding. Moreover, the largest gains are concentrated on two related RadioML2016 benchmarks. These limitations leave uncertainty about how consistently the substantial gains would persist on genuinely new samples or scenarios.

  2. The practical benefit under estimated SNR remains unverified. Section 5.5 studies Gaussian perturbations of true SNR rather than an actual signal-based SNR estimator. Across the tested configurations, mean net gains first become nonpositive at sampled error standard deviations around 3–4 dB, indicating that attainable estimation quality could substantially affect usefulness. Soft routing additionally uses the error scale specified in the simulation, whose determination and possible misspecification in practice are not evaluated. Although these limitations are acknowledged, the study does not establish how much benefit remains when SNR and its error characteristics must be estimated from received signals.

Questions for authors
  1. Which key design choices in the follow-up study were informed by earlier test results? If genuinely unused samples are available, could you confirm the main RadioML2016 gain on a representative backbone, with the backbone, readout-selection procedure and evaluation-region definitions fixed in advance? An evaluation under otherwise matched conditions would help establish whether the large gain persists beyond the retrospective analysis.

  2. Could you evaluate an actual signal-based SNR estimator on a representative RadioML2016.10B backbone and report its errors by SNR range, together with classification gains over raw predictions and SNR-free global readout?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.