Scalable Learning in Recurrent Spiking Neural Networks without Backpropagation

A recent paper on arXiv introduces a multi-layer recurrent spiking neural network (SNN) architecture designed for scalable supervised learning without relying on backpropagation or surrogate gradients. This structure features locally dense recurrent layers enhanced by sparse long-range projections to a readout population. To maintain routing efficiency and hardware scalability, long-range connections remain mostly fixed, while synaptic adaptation employs local plasticity mechanisms. The learning framework integrates output layer population-based winner-take-all teaching signals, fixed random broadcast alignment feedback pathways, and low-dimensional modulatory neurons. This innovative method tackles the issue of scalable learning in deep recurrent SNNs with sparse connections, providing a biologically inspired alternative to backpropagation.