Learning capacity allocation for stable sequential learning in recurrent spiking neural networks

Sequential learning remains challenging due to instability and interference caused by continuous parameter adaptation. Most existing approaches focus on how parameters are updated, which implicitly assumes that adaptation is uniformly permitted across the model. Here we identify learning capacity allocation, defined as the problem of determining where adaptation is permitted, as an upstream and critical factor shaping sequential learning dynamics. We introduce IP²-RSNN, a two-stage framework that allocates intrinsic neuronal plasticity at the task-family level and evaluates its effects during sequential learning. Across multiple cognitively motivated delayed-response task families, task-dependent intrinsic capacity allocation substantially improves learning stability and adaptation efficiency compared to misallocated intrinsic plasticity. By contrast, uniform intrinsic adaptation does not achieve the same stability as task-dependent allocation. Beyond performance, intrinsic capacity allocation induces structured specialization at both neuron and network levels, which is absent under uniform adaptation in continuous recurrent networks. These results reveal that learning capacity allocation is a principled route towards stable and interpretable sequential learning.