Neuromorphic Engineering Algorithms for Edge ML and Spiking Neural Networks

This forum session explores efficient and robust algorithms for training spiking neural networks on edge and neuromorphic hardware. Topics include surrogate gradients, temporal batch normalization, interpretable spike activation maps, homeostatic plasticity, and meta-learning.

Syllabus

Intro
JOHNS HOPKINS UNIVERSITY
Spiking for tinyML
Batch Normalization Through Time (BNTT) for Temporal Learning
BNTT: Energy Efficiency & Robustness
Training SNNs for edge with heterogeneous demands
Spike Activation Map (SAM) for interpretable SNN
Spiking neurons are binary units with timed outputs
End-to-end training is key for artificial neural networks
Solution: Replace the true gradient with a surrogate gradient
Surrogate gradients self-calibrate neuromorphic systems when they can access the analog substrate variables
Fluctuation-driven initialization and bio-inspired homeostatic plasticity ensure optimal initialization
Holomorphic Equilibrium Propagation Computes Exact Gradients Through Finite Size Oscillations
Technical Program Committee