Class-Incremental Learning for Recognition of Complex-Valued Signals

Signal recognition, essential in both military and civilian applications, often deals with an expanding array of signal classes due to the emergence of new communication devices. Current class-incremental learning (CIL) approaches, primarily devised for image-based tasks, prove less efficient when handling complex-valued signals. Moreover, global fine-tuning is not feasible due to its high computational cost. This paper proposes a complex-valued CIL framework, coined as C-SRCIL, engineered to identify complex-valued signals. C-SRCIL features a decoupled feature extractor to limit catastrophic forgetting and updating costs while ensuring the effectiveness of feature representation for CIL with complex-valued neural networks and a carefully designed integrated loss function. During the incremental stage, C-SRCIL modifies the classifier with an adaptive node fusion-based complex-valued CIL adapter, effectively accommodating the increasing signal classes. This paper also proposes an ambiguous boundary indication method for C-SRCIL which solely depends on the weight correlation of the complex-valued classifier to pinpoint the potential ambiguity of signals. Experimental results on benchmark datasets reveal that C-SRCIL outperforms contemporary techniques, highlighting its capacity to expand classification boundaries of previous models with lower overhead. The ambiguous boundary indication method has also been empirically validated, showing its capability to augment predictive information in C-SRCIL.