Towards Hardware-Software Self-Adaptive Acceleration of Spiking Neural Networks on Reconfigurable Digital Hardware

Self-adaptive systems can dynamically and autonomously modify their parameters or structure to adjust their behavior in response to changes in their environment or input. The resulting flexibility and robustness can be a great advantage in various fields like robotics or industrial processes. Here,...

Full description

Saved in:
Bibliographic Details
Main Authors: Pachideh, Brian, Zielke, Christian, Nitzsche, Sven, Becker, Jurgen
Format: Conference Proceeding
Language:English
Subjects:
adaptive systems
FPGA
Hardware
Learning systems
Network topology
Neural networks
neuromorphic
online learning
partial reconfiguration
Power demand
SNN
Software
Software algorithms
STDP
Online Access:Request full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Self-adaptive systems can dynamically and autonomously modify their parameters or structure to adjust their behavior in response to changes in their environment or input. The resulting flexibility and robustness can be a great advantage in various fields like robotics or industrial processes. Here, we demonstrate the feasibility of combining field-programmable gate arrays (FPGAs) and spiking neural networks (SNNs) to create self-adaptive systems capable of dynamic and autonomous modification of their structure in both software and hardware, allowing them to learn and evolve over time. We present a proof-of-concept implementation of an SNN accelerator that can adapt its network topology using dynamic partial reconfiguration and its weight parameters using online learning methods, such as Spike-Timing Dependent Plasticity (STDP) and anti-STDP. We evaluated different configurations of the accelerator in terms of resource utilization, power consumption, and runtime and tested its online learning capabilities in a classification task using the N-MNIST dataset. In our experiments, the accelerator saved 33% dynamic power by adapting to changing task conditions, and achieved a speedup factor of 5.3 with respect to the input data time resolution. Our findings suggest that combining FPGAs and SNNs can lead to the effective instantiation of SNN hardware resources to meet changing task conditions, thus paving the way for future research on self-adaptive SNNs that use diverse online learning algorithms to improve themselves both in software and hardware.
ISSN:2164-1706
DOI:10.1109/SOCC58585.2023.10257066