FPGA-based fast bin-ratio spiking ensemble network for radioisotope identification

Area of Science:

• Artificial Intelligence
• Nuclear Science & Engineering
• Computer Engineering

Background:

• Spiking neural networks (SNNs) offer energy-efficient computation.
• Radioisotope identification is crucial for nuclear security and research.
• FPGA implementation allows for hardware acceleration of AI models.

Purpose of the Study:

• To develop and implement an FPGA-based bin-ratio ensemble SNN for radioisotope identification.
• To optimize the SNN for high accuracy and low power consumption.
• To evaluate the performance of the implemented SNN in terms of accuracy, speed, and energy efficiency.

Main Methods:

• Training and conversion of an ensemble SNN comprising 20 3-layer networks with 1160 neurons.
• Application of learned step quantization (LSQ) and pruning techniques for network compression.
• Implementation on an Artix-7 FPGA board with a 100 MHz clock frequency.

Main Results:

• Achieved 97.04% accuracy in radioisotope identification with less than 1% accuracy loss.
• Network parameter compression reduced size to 30% of original.
• Inference time of 334 μs per sample.
• Estimated power consumption of 157 μJ per inference.

Conclusions:

• The developed FPGA-based SNN demonstrates a viable and efficient solution for radioisotope identification.
• LSQ and pruning are effective techniques for optimizing SNNs for hardware deployment.
• The implementation showcases the potential of SNNs for low-power, high-accuracy edge computing in nuclear applications.