Implementing spiking neural networks for real-time signal-processing and control applications: a model-validated FPGA...

Area of Science:

• Neuroscience
• Robotics
• Computer Engineering

Background:

• Hardware architectures are crucial for simulating large-scale neural networks.
• Existing systems face challenges in real-time performance for closed-loop robotic control.
• Biologically plausible neural networks (NNs) offer advanced control capabilities.

Purpose of the Study:

• To present two FPGA-based hardware architectures for modeling leaky-integrate-and-fire (LIF) neuron networks.
• To enhance performance for real-time applications in mobile robotic vehicles.
• To facilitate the integration of bio-inspired neural processing into real-world control systems.

Main Methods:

• Implementation of two fixed-point arithmetic hardware architectures on a single FPGA.
• Simulation of neural networks comprising over 1000 neurons using biologically plausible models.
• Development of a system for porting floating-point models to fixed-point FPGA representations.

Main Results:

• Successful simulation of large neural networks (>1000 neurons) with biologically plausible models.
• Demonstrated real-time performance suitable for closed-loop robotic control systems.
• Validation of the hardware architecture through a development system and the Whiskerbot project.

Conclusions:

• The developed FPGA architectures effectively model large LIF neural networks.
• The neuroprocessor is suitable for mobile robotic applications requiring real-time neural processing.
• The development system aids collaboration between neuroscientists and engineers for embodied systems.