Class-Incremental Learning for Baseband Modulation Classification: A Comparison
Charles Montes, Todd Morehouse, Ruolin Zhou · 2024
This paper presents a comprehensive study on the capabilities of class-incremental learning in the context of baseband modulation classification. Despite the growing interest in incremental learning, there is a lack of information specifically addressing its application in the radio frequency (RF) domain for modulation classification. This study aims to fill this gap by investigating the effects of incremental learning when applied to different increments of classes, methods, and types of methods like exemplars. We explore various methods including non-incremental learning, cross-entropy distillation, and bias correction, and evaluate their performance in the context of incremental learning. Multiple incremental scenarios are considered including adjusting the step size of the number of classes learned, and adjusting the number of exemplars used during incremental training. The capabilities of incremental learning are gauged based on their ability to continually learn new classes without forgetting the previously learned ones. The evaluation is performed on the DeepSig 2018A dataset, which comprises 24 classes, providing a robust platform to assess the capabilities of incremental learning. The results of this study provide valuable insights into the potential of incremental learning in baseband modulation classification.