Deep Learning with Width-Wise Early Exiting and Rejection for Computational Efficient and Trustworthy Modulation Classification

The development of trustworthy and efficient Deep Learning (DL) models is vital for wireless communications, supporting tasks such as automatic modulation classification (AMC), spectrum use, and network optimization. Yet, deploying DL on resource-constrained edge devices remains challenging due to energy and reliability concerns. We propose a width-wise early exiting architecture, a variation of conventional early exiting that enables classification after processing only part of a signal frame. To further improve reliability, we introduce an early rejection mechanism, applying confidence-based abstention both at intermediate exits and the final output. In AMC experiments, our model achieves on average 40% less computation (up to 60% in some cases), while improving classification accuracy by 3% in low-SNR conditions. These results highlight the potential of our approach for robust, efficient, and trustworthy ML deployment in wireless environments.