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

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.