Related Experiment Videos
Compact hardware liquid state machines on FPGA for real-time speech recognition
Benjamin Schrauwen1, Michiel D'Haene, David Verstraeten
1Department of Electronics and Information Systems, Ghent University, Sint-Pietersnieuwstraat 41, 9000 Gent, Belgium. Benjamin.Schrauwen@UGent.be
Summary
This study explores hardware for Spiking Neural Networks (SNNs), focusing on real-time speech recognition. A novel, compact, and scalable SNN architecture is presented for efficient, real-time processing.
Area of Science:
- Neuromorphic Engineering
- Computer Engineering
- Artificial Intelligence
Background:
- Spiking Neural Networks (SNNs) are increasingly implemented in hardware due to their efficiency and performance advantages over traditional neural networks.
- Hardware implementations are crucial for deploying SNNs in real-time applications.
Purpose of the Study:
- To explore the hardware design space for Spiking Neural Networks (SNNs) for application-driven digital hardware.
- To implement real-time, isolated digit speech recognition using a Liquid State Machine (LSM) on digital hardware.
Main Methods:
- Testing and improving existing hardware architectures for LSMs.
- Developing a novel, scalable, and serialized digital hardware architecture for SNNs.
- Implementing leaky integrate-and-fire neuron models with exponential synaptic dynamics.
Main Results:
- Existing SNN hardware architectures were found to be too fast and area-consuming for the target speech recognition application.
- A new scalable, serialized SNN architecture was developed, achieving real-time processing with a compact footprint.
- The developed architecture supports leaky integrate-and-fire membranes and exponential synaptic models.
Conclusions:
- There is a significant unexplored hardware design space for Spiking Neural Networks.
- The presented scalable, serialized architecture offers an efficient solution for real-time SNN applications like speech recognition.