Reproduce “An error-propagation spiking neural network compatible with neuromorphic processors” on OpenResearch

About this paper

Matteo Cartiglia, Germain Haessig, Giacomo Indiveri

Researchers at the Institute of Neuroinformatics, University of Zurich and ETH Zurich, developed a spiking neural network (SNN) architecture that approximates the back-propagation algorithm using local, spike-based error propagation. This SNN is compatible with mixed-signal neuromorphic processors, enabling on-chip learning for spatio-temporal pattern recognition and discrimination tasks with high specificity.

Key insights

• Local, spike-based error propagation mechanisms can effectively approximate global gradient-based learning rules in multi-layer SNNs, making them suitable for on-chip learning on power-constrained neuromorphic hardware.
• The multi-compartment neuron model, especially the apical compartment, is crucial for integrating top-down feedback and lateral inhibition to minimize prediction errors, enabling adaptive learning.
• Specific local plasticity rules, incorporating factors like membrane current differences and hysteresis, can enable stable and robust weight updates despite the noise and imprecision inherent in analog neuromorphic circuits.

Summary by alphaXiv.