Related Experiment Video
Updated: Aug 17, 2025
Closed-loop Neuro-robotic Experiments to Test Computational Properties of Neuronal Networks
Published on: March 2, 2015
E-prop on SpiNNaker 2: Exploring online learning in spiking RNNs on neuromorphic hardware
Amirhossein Rostami1, Bernhard Vogginger1, Yexin Yan1
1Chair of Highly-Parallel VLSI-Systems and Neuro-Microelectronics, Faculty of Electrical and Computer Engineering, Institute of Principles of Electrical and Electronic Engineering, Technische Universität Dresden, Dresden, Germany.
This study demonstrates efficient, low-memory training of Spiking Recurrent Neural Networks (SRNNs) at the edge using the E-prop algorithm on SpiNNaker 2. This approach significantly reduces energy consumption compared to GPUs for real-time keyword spotting.
Area of Science:
- Neuromorphic computing
- Edge artificial intelligence (AI)
- Deep learning
Background:
- Deep learning models are increasingly deployed at the edge, necessitating efficient training methods.
- Traditional training of recurrent neural networks (RNNs) requires significant memory for Back Propagation Through Time (BPTT).
- Spiking Recurrent Neural Networks (SRNNs) offer a biologically inspired alternative, with E-prop addressing BPTT's memory limitations.
Purpose of the Study:
- To implement and evaluate the E-prop algorithm for training SRNNs on the SpiNNaker 2 neuromorphic system.
- To develop a parallelization strategy for efficient memory and compute utilization on SpiNNaker 2.
- To assess the feasibility of real-time, on-device training for applications like keyword spotting.
Main Methods:
- Implementation of the E-prop algorithm on the SpiNNaker 2 prototype.
- Development of a parallelization strategy to distribute network training across SpiNNaker 2's ARM cores.
- Training of an SRNN from scratch on the Google Speech Command dataset for keyword spotting.
Main Results:
- Achieved 91.12% accuracy in keyword spotting with only 680 KB of memory for training.
- Demonstrated significant memory efficiency compared to other spiking neural networks.
- Profiling indicated E-prop's suitability for edge training and potential for architectural optimization on SpiNNaker 2.
Conclusions:
- The E-prop algorithm enables efficient, low-memory SRNN training on neuromorphic hardware like SpiNNaker 2.
- This approach is highly competitive for edge AI applications, offering substantial energy savings over GPUs.
- Further optimization of SpiNNaker 2 architecture can accelerate online learning at the edge.
