Close

Presentation

SpikeDyn: A Framework for Energy-Efficient Spiking Neural Networks with Continual and Unsupervised Learning Capabilities in Dynamic Environments
Time
Location
Event Type
Research Manuscript
Virtual Programs
Hosted in Virtual Platform
Keywords
AI/ML System Design
Topics
Design
DescriptionSpiking neural networks bear the potential of unsupervised and continual learning capabilities, but they consume high energy, which makes them difficult to be deployed in the energy-constrained scenarios. Towards this, we propose SpikeDyn framework that employs 1) reduction of neuronal operations, 2) model size search algorithm, and 3) a continual and unsupervised learning algorithm through adaptive learning rates, threshold potential, and weight decay. SpikeDyn reduces the energy, on average by 51% for training and by 37% for inference, while improving the accuracy, on average by 21% for the most recently learned task and by 8% for the previously learned tasks.