Excitation–Inhibition Balance Facilitates Meta-Learning in Spiking Neural Networks for Few-Shot Rapid Adaptation

Meta-learning methods are an effective approach to tackling the challenges of few-shot learning. However, current meta-learning approaches based on artificial neural networks do not introduce brain-inspired learning mechanisms, thereby hindering these models from achieving few-shot learning efficiencies and flexibilities comparable to human performance. Furthermore, existing brain-inspired spiking neural network approaches for meta-learning fail to meet Dale’s law constraints on excitation–inhibition balance during synaptic plasticity-driven weight regulation. This limitation compromises both the biological plausibility and the learning capability of spiking meta-learning. This work introduces a Dale’s law-constrained neural excitation–inhibition balance mechanism for meta-learning the initialization of spiking neural networks, constructing internal representations broadly applicable to many few-shot tasks. Specifically, to ensure strict compliance with Dale’s law (neurotransmitter type invariance) during synaptic weight updates, we integrate a modified projected gradient descent method with synaptic plasticity, thereby maintaining the invariance of neural excitatory and inhibitory properties. Building upon this, we incorporate the Dale’s law-constrained excitation–inhibition balance mechanism into spiking meta-learning and propose the excitation–inhibition balance-based spiking meta-initialization (EI-SMI) algorithm. Next, to enhance the stability and generalization capability of spiking meta-initialization, we employ a multistep weighted loss strategy to guide the network’s meta-learning process. Extensive experiments on five benchmark datasets (Omniglot, Mini-ImageNet, CUB, CIFAR-FS, and CWRU) demonstrate that the proposed EI-SMI outperforms early artificial neural network-based meta-learning approaches on certain specialized tasks and surpasses Spiking MAML across all tasks. The experimental results indicate that our method serves as a solid empirical foundation for developing meta-learning algorithms that are both biologically plausible and mechanistically credible.