acceptodds
Withdrawn from ICLR 2027

BETA-AMR: Batch-Consistent Efficient Temporal Modeling for Automatic Modulation Recognition

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Automatic modulation recognition (AMR) plays a vital role in wireless communications by identifying signal modulation types from raw IQ samples under strict efficiency constraints. Existing methods focus almost exclusively on deployment metrics like parameter size and inference latency, while accuracy may change substantially when the training batch size changes. To study this trade-off, we propose BETA-AMR, a batch-consistent, highly efficient temporal architecture designed for parallel training and efficient single-sample deployment. Rather than relying on sequential state propagation or heavy multi-head attention, BETA-AMR introduces a fully parallel, dilated temporal-convolution backbone that scales the receptive field without sequential O(T) dependencies. Group Normalization avoids batch-dependent statistics in the feedforward path, while experiments separately measure sensitivity to training batch size. Furthermore, a lightweight, non-pairwise temporal aggregation layer adaptively isolates high-information burst regions with minimal computational overhead compared to O(T^2C) self-attention. Under a unified protocol, BETA-AMR achieves the highest accuracy among compared methods under the batch-2048 protocol across three length-128 benchmarks with an ultra-low batch-1 latency of 1.04 ms. When scaling training batch size from 128 to 2048, its accuracy drops by only 0.75±0.14 points on RML2016.10A and 0.80±0.35 points on RML22.01A across multiple seeds. On the longer RML2018 benchmark, expanding its receptive field boosts accuracy by 1.7–1.8 points, bridging the gap to a similarly sized recurrent hybrid to under 0.1 points. Together these results motivate reporting training-batch sensitivity and execution speed alongside accuracy and deployment latency.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
automatic modulation recognition, temporal modeling, efficient deep learning, batch-consistent training, streaming inference, wireless signal classification

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_fT55OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes BETA-AMR, an efficient architecture for automatic modulation recognition that combines amplitude–phase input representations, a parallel dilated temporal convolutional encoder, Group Normalization, and lightweight temporal attention pooling. The method is evaluated in terms of recognition accuracy, inference efficiency, and sensitivity to training batch size. Experiments on four benchmarks show that the compact model achieves the highest accuracy among the evaluated baselines on three short-sequence datasets under the batch-2048 training protocol, while a larger variant approaches the recurrent baseline on longer sequences.

Critical strengths
  1. The compact model achieves competitive accuracy with low single-sample inference latency.

  2. The evaluation covers four benchmarks and examines sensitivity to training batch size.

  3. Component ablations and multi-seed experiments help characterize the model's behavior.

Critical weaknesses
  1. Limited methodological novelty. The architecture primarily combines established components: dilated temporal convolutions, GroupNorm, and attention pooling. Direct comparisons with well-tuned TCN or ModernTCN baselines are missing, making the contribution beyond existing convolutional architectures unclear.

  2. Insufficient support for the batch-consistency motivation. In Table 5, the GN model loses 0.60 percentage points when increasing the training batch size from 128 to 2048, whereas the BN variant loses only 0.03 points. Although the authors acknowledge this result, the role of GN in motivating a “batch-consistent” architecture remains unclear.

  3. Inconsistent benefits from learned temporal pooling. In Table 5, learned pooling improves accuracy over average pooling by only 0.12 percentage points at batch 2048 and performs worse at batches 128 and 1024. These results provide limited support for presenting this module as a key design contribution.

  4. Insufficient control of training-budget effects. The main paper does not clearly establish how optimization budgets are controlled across training batch sizes. Without matched-compute or time-to-accuracy comparisons, it is difficult to determine whether the observed accuracy retention translates into a practical training-efficiency advantage.

Questions for authors
  1. The method combines established components without direct comparisons against closely related TCN baselines, leaving its methodological contribution unclear.

  2. The source of training-batch robustness remains unclear, as BN shows a smaller accuracy drop and optimization-budget effects are insufficiently addressed.

  3. Using AP inputs for all baselines may disadvantage methods designed around I/Q-specific operations, raising concerns about comparison fairness.

  4. Learned temporal pooling provides inconsistent gains over average pooling, weakening its justification as a key design contribution.

Reviewer_Jw6JOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes BETA-AMR, a lightweight automatic modulation recognition model that balances recognition accuracy, batch-consistent, and low inference latency. Specifically, BETA-AMR employs a fully parallel dilated temporal-convolution backbone to expand the receptive field without sequential dependencies, combined with Group Normalization to avoid batch-dependent statistics. A lightweight, non-pairwise temporal aggregation layer emphasizes informative burst regions. Extensive experiments have demonstrated the effectiveness of the proposed method.

Critical strengths
  • The study focuses on the ability to maintain performance after changes in training batch size, which is relatively novel.
  • The paper is easy to follow.
  • The ablation experiment is performed thoroughly.
Critical weaknesses
  • Table 3 is too large and looks strange; I suggest revising it.
  • The latest model is TCN-GRU, proposed in 2024. Can it be compared with newer models?
  • The main contribution of this paper is more of a combination of existing technologies (such as GroupNorm), and its unique value compared to existing methods needs to be further explained.
  • Since different training seeds can affect model accuracy, it is best to average the results from multiple training iterations using each method to overcome the interference of randomness.
Questions for authors
  • In Figure 3, the minimum batch size is 128. Will there be any impact if the batch size is too small?
Reviewer_rvfcOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces BETA-AMR, a batch-consistent and efficient temporal architecture designed to support both parallel training and high-efficiency, single-sample deployment. Moving away from designs that rely on sequential state propagation or complex multi-head attention mechanisms, BETA-AMR employs a fully parallel backbone based on dilated temporal convolutions—an approach that effectively expands the receptive field without requiring sequential O(T)O(T) dependencies.

Critical strengths
  1. BETA-AMR gets good results on three short-signal datasets and processes each signal quickly.
  2. Repeated tests show that its accuracy stays fairly stable when the training batch size changes.
Critical weaknesses
  1. While the paper emphasizes efficient single-sample deployment, the latency figures, measured using compiled BF16 inference on an RTX 4090, it fails to demonstrate the model's advantages on resource-constrained receiving devices. The study lacks evaluations of latency, memory usage, and energy consumption on representative deployment hardware, as well as validation using signals captured in real-world scenarios

2.Dilated temporal convolutions, GroupNorm, and temporal weighted pooling are all well-established techniques.

The paper needs to articulate more clearly where its specific contributions lie and whether these components yield a synergistic effect that goes beyond the sum of their individual benefits

Questions for authors
  1. Does BETA-AMR remain efficient on resource-constrained devices and accurate on real-world signals?

  2. What is novel beyond combining existing components, and is their synergy demonstrated?

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

Related papers

Discussion (0)

Sign in to comment.