Accelerating Transformer Architectures for Automatic Modulation Classification via Gated Activation Patch Selection

This paper presents a novel approach to Automatic Modulation Classification (AMC) that leverages a Gated Activation Patch Selection (GAPS) mechanism to enhance the efficiency of Transformer architectures. Our method addresses the computational challenges associated with Transformer-based AMC models by filtering out less informative signal patches before processing. The GAPS module utilizes convolutional layers to extract key features from the input signal and quantifies the importance of each patch through adaptive aggregation. By employing a Transformer encoder, the selected critical patches are transformed into embeddings, facilitating self-attention computations. Extensive experiments demonstrate that our method not only maintains competitive classification performance but also significantly reduces computational complexity, resulting in faster training and lower inference latency. Notably, our method achieves a 17.94% FLOPs reduction without any loss in performance, along with a similar degree of training acceleration, while the model's accuracy decreases by only 1.50% when the input signal length is reduced by 50%.