Training the Untrainable - Learning to Backpropagate Through Spikes and Time

Learn to overcome the fundamental challenges of training spiking neural networks through innovative backpropagation techniques in this conference talk that explores how to make traditionally "untrainable" spike-based systems amenable to gradient-based learning. Discover the mathematical foundations and practical approaches for propagating gradients through discrete spike events and temporal dynamics, examining why conventional backpropagation fails in spiking networks and how novel methods enable effective training. Explore the intersection of neuroscience-inspired computing and machine learning optimization, understanding how to bridge the gap between biological plausibility and computational efficiency in neuromorphic systems. Gain insights into the latest research developments that make it possible to train spiking neural networks for complex tasks while maintaining their inherent advantages of energy efficiency and temporal processing capabilities.