NeuMoSync Enhances Continual Learning with Brain-Inspired Neuromodulation.

Continual learning systems face a significant challenge: learning new tasks sequentially without forgetting previously acquired knowledge, a problem known as catastrophic forgetting. Deep neural networks often struggle with this, exhibiting plasticity loss and poor knowledge transfer. Drawing inspiration from the brain's global neuromodulatory mechanisms, a new architecture called NeuMoSync aims to address these limitations. NeuMoSync extends standard neural networks by incorporating learnable feature vectors for each neuron, which track the network's historical context. A higher-level module then synthesizes neuron-specific signals, conditioned on both current inputs and the network's evolving state. These signals adaptively regulate activation dynamics and synaptic plasticity, allowing the network to maintain flexibility while integrating new information. Evaluated across various continual learning benchmarks, including memorization, concept drift, class-incremental, and domain-incremental learning, NeuMoSync consistently outperforms existing methods in retaining plasticity and improving both forward and backward adaptation. Ablation studies confirm the necessity of each component, and analysis of the learned signals reveals interpretable coordination patterns, highlighting the potential of brain-inspired global coordination for robust, adaptive continual learning.