Continual Learning: Improved Efficiency and Class Dependency Examination with SHAPC
Year: 2026
Type: conference-paper
Abstract: Continual learning (CL) enables artificial intelligence (AI) systems to adapt to new information while mitigating catastrophic forgetting (the loss of previously acquired knowledge). The popularity of these algorithms in safety-critical domains has necessitated interpretability methods to understand model predictions. One such method derived from Shapley Additive Explanations (SHAP) is SHAP Consistency (SHAPC). Its global variant, SHAPC-Mean, quantifies the decision attribution stability throughout the learning process. The existing SHAPC-Mean requires iterations over all samples and task-pair combinations, making it computationally expensive. We propose three optimized approaches to computing SHAPC-Mean that reduce samples, task-pair iterations, or a combination of both. This decreases overall computation time while maintaining metric performance. Additionally, we introduce a classwise SHAPC-Mean metric to assess potential class-dependent biases. We discover SHAPC-Mean to be class-independent for most algorithms.
Institutions: Rowan University, Fairleigh Dickinson University
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7212295271