Spiking Neural Network Based Low-Power Radioisotope Identification using FPGA

Abstract

this paper presents detailed methodology of a Spiking Neural Network (SNN) based low-power design for radioisotope identification. A low power cost of 72 mW has been achieved on FPGA with the inference accuracy of 100% at 10 cm test distance and 97% at 25 cm. The design verification and chip validation methods are presented. It also discusses SNN simulation on SpiNNaker for rapid prototyping and various considerations specific to the application such as test distance, integration time and SNN hyperparameter selections.

An FPGA Implementation of Convolutional Spiking Neural Networks for Radioisotope Identification

“…Table I shows the comparison between this convolutional structure and the fully-connected structure in [5]. The CSNN in this work supports more than ten times higher data dimensions with less than half of the weights.…”

“…Average pooling was chosen because of the ease with which it can be implemented in a rate-based spiking model [6] and its comparable performance with "max" pooling. The fully-connected output layer is implemented also in TDM manner based on the microarchitecture in [5]. weight…”

“…Table I shows the worst case maximum operations per each input spike event. Compared with the fully-connected structure in [5], the convolutional structure saves half of the operations, which leads to potential dynamic power savings.…”

“…Furthermore, those algorithms are normally implemented on power-hungry computing architecture such as CPU or GPU, which are not always the most suitable for power constrained edge devices. In [4] [5], this unnecessary power overhead of the frame-based methods for radioisotope ID was pointed out and an event-based Spiking Neural Network (SNN) architecture was proposed and implemented on a Field-Programmable Gate Array (FPGA). In an event-based manner, the always-on energy histogram generation process is removed and processing only takes place when detection events happen.…”

“…In [5] a proof-of-principle was presented, in which a shallow fully-connected SNN architecture demonstrated the high power efficiency and high accuracy on the radioisotope ID task by using an event-based SNN. However, the input to this model was resampled to only cover 100 energy channels in the input layer, limiting the system's application in cases where isotopes have similar histograms or where input has a low signal-to-noise ratio.…”