Adaptive Knowledge Consolidation for Sequential Learning in Deep Neural Networks
Abstract: Catastrophic forgetting is a major challenge in sequential deep learning, where learning new tasks can degrade previously acquired knowledge, particularly in evolving healthcare applications. Existing approaches mainly rely on replay, regularization, or architectural strategies, while the integration of adaptive, biologically inspired sleep mechanisms in conventional deep networks remains limited. To address this gap, this paper proposes a conceptual Adaptive Knowledge Consolidation (AKC) framework integrating wake-phase learning, NREM-inspired replay, REM-inspired knowledge integration, synaptic homeostasis, and adaptive sleep scheduling. The framework is designed for sequential image classification, with planned validation using Alzheimer?s disease-related brain MRI data. Evaluation will consider accuracy, forgetting measure, backward and forward transfer, sensitivity, specificity, F1-score, and AUROC. As implementation is ongoing, no numerical results are claimed. The study establishes the framework, mathematical formulation, and experimental protocol for future empirical validation.
Keywords: continual learning; catastrophic forgetting; knowledge consolidation; sleep-inspired learning; deep neural networks; replay; medical imaging; Alzheimer's disease
How to Cite?: Bidura V M, Vasanthakumar S, "Adaptive Knowledge Consolidation for Sequential Learning in Deep Neural Networks ", Volume 15 Issue 10, October 2026, International Journal of Science and Research (IJSR), Pages: 518-525, https://www.ijsr.net/getabstract.php?paperid=SR261005100507, DOI: https://dx.doi.org/10.21275/SR261005100507