Acceleration of spiking neural network based pattern recognition on NVIDIA graphics processors.

Area of Science:

• Computational Neuroscience
• Artificial Intelligence
• High-Performance Computing

Background:

• The research community is developing large-scale, biologically plausible neural network models for vision.
• These systems require computationally intensive neuron models, such as Izhikevich and Hodgkin-Huxley, which are more accurate than simpler models.
• Existing computational resources limit the scale and complexity of these simulations.

Purpose of the Study:

• To investigate the feasibility of using graphics processing units (GPUs) to accelerate spiking neural network (SNN) based character recognition.
• To evaluate the performance of different NVIDIA general-purpose GPU (GPGPU) platforms for accelerating complex neuron models.
• To compare GPU acceleration against highly optimized central processing unit (CPU) implementations.

Main Methods:

• Implemented two SNN character recognition network versions using Izhikevich and Hodgkin-Huxley neuron models.
• Tested performance on three NVIDIA GPGPU platforms: GeForce 9800 GX2, Tesla C1060, and Tesla S1070.
• Compared GPGPU performance against a quadcore 2.67 GHz Xeon CPU with optimized multi-core and vector processing.

Main Results:

• GPGPUs demonstrated substantial speedups compared to conventional CPUs for both Izhikevich and Hodgkin-Huxley models.
• The Tesla S1070 achieved speedups of 5.6x for the Izhikevich model and 84.4x for the Hodgkin-Huxley model over the optimized CPU.
• The CPU implementation leveraged all four cores and vector data parallelism.

Conclusions:

• GPUs are highly effective for accelerating computationally demanding neural network simulations, particularly those using complex neuron models.
• GPGPU acceleration is a viable strategy for enabling large-scale, biologically realistic neural vision systems.
• The findings support the suitability of GPUs for advancing research in computational neuroscience and AI.