Reproduce “A Hybrid ANN-SNN Architecture for Low-Power and Low-Latency Visual Perception” on OpenResearch

About this paper

Asude Aydin, Mathias Gehrig, Daniel Gehrig, Davide Scaramuzza

Researchers at the Robotics and Perception Group at the University of Zurich developed a hybrid Artificial Neural Network (ANN) and Spiking Neural Network (SNN) architecture that addresses SNN's slow convergence and state decay by periodically initializing SNN internal states from the ANN. This approach enables 2D and 3D human pose estimation with an 8x reduction in power consumption compared to pure ANNs while maintaining high accuracy, and is 26.7 times more energy-efficient than state-of-the-art event-based ANNs.

Key insights

• Periodically initializing SNN membrane potentials with ANN features significantly accelerates SNN convergence, eliminating warm-up periods and mitigating state decay.
• Ending state initialization modules with Batch Normalization allows for 'meta-stable' SNN neurons (initialized near firing threshold), leading to immediate responsiveness and reduced spiking activity overall.
• This hybrid approach enables the SNN to maintain high accuracy and low latency, while considerably reducing total energy consumption compared to pure SNNs (by preventing wasted energy on convergence) and pure ANNs.

Summary by alphaXiv.