Radioisotope identification with neuromorphic methodology: different solutions and evaluations

Abstract

Early detection of radioisotopes plays an increasingly important role in the modern world. It allows the possibility of quick countermeasures when faced with potentially hazardous radioactive materials like dirty bombs, and nuclear leakage. This could secure the lives of the innocent in populated areas including airports, stadiums or ports. A light-weight compact handheld device could be used in this situation for the patrol team. However, the operating hours for these devices are normally constrained by the batteries they carry. More efficient al- gorithms or solutions are needed for this resource-constraint application to extend the battery life so that security patrol is not frequently interrupted by the recharge.

Event-based processing is a novel technique that allows the computing unit to operate only when there is a key event while staying idle otherwise. Spiking neural network (SNN) is a promising candidate for event-based processing and also known as neuromorphic method- ology due to the biomimicry plausibility, which could be easily implemented and still offer comparable accuracy to its counterpart — artificial neural network (ANN), which is notoriously power-hungry.

In this research work, it will be demonstrated that using SNN for radioisotope identification (RIID) is possible and capable of achieving the same or even better accuracy when compared with ANNs. Meanwhile, the power consumption of the proposed method on a field program- mable gate array (FPGA) shows that power reduction is highly significant compared with the old software implementation on a smartphone.

The task has been delivered in two parts, we first attempted an unsupervised Spike-Timing- Dependent Plasticity (STDP) SNN implementation on SpiNNaker, an emulation platform for SNN. This demonstrates the capability of classifying radioisotopes using purely SNN compat- ible training methods and architecture. We then managed to implement a more complex bin-ratio ensemble SNN (BESNN) on FPGA with better performance. To achieve this implementation, a new SNN conversion method was created to facilitate the digital hardware implementation. This conversion flow allows the highly sparse weight matrix representation without sacrificing overall accuracy. In the meantime, the power consumption of the mentioned design has been characterised, which could be used to estimate the battery life of a handheld system while functioning.

Even though this design has been validated on an FPGA, further squeeze for the power saving is possible if an application specific integrated circuits (ASIC) could be delivered. Furthermore, the analogue unit used in the design is a compromise given that the logarithm could not be done by a spiking neuron at the moment. This prevents an end-to-end application, which is preferred for higher integration and potentially more power conservation.

According to our knowledge, applying neuromorphic methodology to address RIID represents uncharted territory, especially in the context of power characterisation, an aspect that has not been explored previously. This research work fills the gap that is present in the research field and also offers a functional low-power prototype for the handheld RIID device producer.

This project pioneers the use of an event-based processing algorithm for radioisotope identi- fication, marking a significant advancement in the field. Leveraging Spiking Neural Networks (SNNs) on specialised hardware, the project establishes a comprehensive application flow, showcasing the efficacy and potential of SNNs in this domain.

The implementation of an unsupervised STDP algorithm for radioisotope identification is also groundbreaking, introducing a local self-learning rule for complex tasks beyond handwritten digit recognition.

Additionally, the bin-ratio ensemble project achieves remarkable accuracy, setting new bench- marks in the field. It represents the first ensemble SNN application in radioisotope identifica- tion, further enhanced by an innovative ANN-SNN conversion method with iterative pruning to reduce computational overhead.

Furthermore, this research provides detailed insights into sparse SNN construction and char- acterises hardware implementation, shedding light on power and energy consumption con- siderations.