Reproduce “ETLP: Event-based Three-factor Local Plasticity for online learning with neuromorphic hardware” on OpenResearch

About this paper

Fernando M. Quintana, Fernando Perez-Peña, Pedro L. Galindo, Emre O. Neftci, Elisabetta Chicca, Lyes Khacef

Researchers at the University of Cádiz, Forschungszentrum Jülich, and the University of Groningen developed Event-based Three-factor Local Plasticity (ETLP), an online learning rule for Spiking Neural Networks that is fully local in space and time. This rule leverages Direct Random Target Projection and adaptive thresholds to enable energy-efficient, on-chip learning on neuromorphic hardware, demonstrating competitive accuracy on N-MNIST and SHD datasets while showing low computational complexity and resource utilization in an FPGA proof-of-concept.

Key insights

• Adaptive thresholds in ALIF neurons are crucial for enhancing accuracy and regulating firing rates in online, local learning rules like ETLP and eProp, but are less effective or detrimental for offline gradient-based methods like BPTT.
• Recurrent connections significantly improve accuracy for all learning rules on spatio-temporal datasets, highlighting their importance in processing sequential information within SNNs.
• ETLP's inherent locality translates to O(1) space and O(pNM) time complexity, offering substantial efficiency advantages over BPTT and enabling direct, resource-efficient implementation on neuromorphic hardware platforms.

Summary by alphaXiv.