Toward Balanced Continual Learning via Fine-Grained Neuronal Intervention Inspired by Memory Consolidation

ICLR 2026 Conference Submission5119 Authors

Keywords: Continual Learning; Stability-Plasticity Dilemma

Abstract:

Continual learning confronts the fundamental stability-plasticity dilemma between preserving previously acquired knowledge and adapting to novel tasks. Existing approaches employ coarse-grained network-level regularization that fails to capture the fine-grained neuronal dynamics essential for effective stability-plasticity orchestration. The human brain resolves this challenge through memory reconsolidation ---a neural mechanism that selectively reactivates task-relevant memory traces during retrieval, temporarily destabilizing them to enable integration of new information while preserving task-irrelevant memories. Inspired by this neurobiological principle, we introduce K-RECON, a neuron-level continual learning architecture that operationalizes memory reconsolidation through fine-grained neural pathway modulation. Our approach orchestrates stability and plasticity via two complementary components: i.Selective Reactivation Module that performs controlled reactivation and consolidation blockade of task-relevant neuronal clusters and memory amalgamation, and ii. an Adaptive Consolidation Module that enforces parameter protection for inactive neuronal clusters while strategically releasing connections from obsolete tasks. This neuron-level intervention is theoretically grounded within a unified optimization framework, enabling seamless integration into existing continual learning paradigms as a plug-and-play enhancement. Extensive evaluation across diverse continual learning benchmarks validates K-RECON's effectiveness as a model-agnostic architectural enhancement. Notably, on CIFAR-100 sequential classification tasks, our framework achieves a remarkable 6.43% improvement in average incremental accuracy relative to EWC, establishing neuron-level memory reconsolidation as an effective technique for continual learning. Code for experiments is
available at https://anonymous.4open.science/r/K_Recon11-CF57

Primary Area: transfer learning, meta learning, and lifelong learning

Submission Number: 5119