Synchronized Membrane Potential-driven Plasticity for Spiking Neural Networks

Spiking neural networks (SNNs) have gained increasing attention due to low power consumption and potential computing capability. Aggregate-label (AL) learning is a biologically inspired algorithm that guides neurons to fire in desired patterns, which enables SNNs to produce specific spike counts in response to given inputs. However, both threshold-driven and membrane potential-driven AL algorithms suffer from inefficient computation and unsatisfactory learning performance, which limits their usability in practical applications. To overcome these disadvantages, this article proposes a novel synchronized membrane potential-driven plasticity (SMDP) algorithm. Inspired by biological burst-firing phenomena, SMDP assigns spiking coefficients to each spike event, allowing neurons to emit multiple spikes simultaneously. Furthermore, it adjusts multiple spikes in a single learning step, significantly improving training performance and reducing runtime. Compared with state-of-the-art learning algorithms, experimental results verify the effectiveness and efficiency of multilayer SMDP on both temporal credit-assignment problems and complicated machine-learning tasks. These results credibly demonstrate that the multilayer SMDP represents significant progress in constructing “brain-like” SNNs with larger network architectures, accelerated training process, and improved performance. The source code of SMDP is available at https://github.com/SZU-ADDG/SMDP.