Efficient Neutron/Gamma Classification ML Models for eFPGA Implementation

Core Concepts

Feasibility study of deploying simple BDT and fcNN ML models on eFPGAs for neutron/gamma classification, with a focus on resource efficiency.

Abstract

The study explores the parameter space for implementing boosted decision tree (BDT) and fully-connected neural network (fcNN) models on an eFPGA fabric, with a focus on resource efficiency. The task is neutron/gamma classification using data collected from an AmBe sealed source incident on Stilbene, optically coupled to an OnSemi J-series SiPM.
Key highlights:

Investigated relevant input features and the effects of bit-resolution and sampling rate, as well as trade-offs in hyperparameters for both ML architectures, while tracking total resource usage.
Compared the performance of ML models to the standard charge comparison pulse shape discrimination (PSD) method, and found that an 8-bit ADC with 15 MS/s sampling rate can maintain high neutron efficiency at low gamma leakage.
For BDT models, a max depth of 3 and 50 boosting rounds were found to be a good balance between performance and resource usage. Flip-flop usage was less than 3% and LUT usage less than 25% of the target Artix 7 FPGA.
For fcNN models, a single hidden layer of size 16 with 8-bit weights/biases and 12-bit activations, along with 30% sparsity, achieved high neutron efficiency. Flip-flop usage was less than 3% and LUT usage less than 10% of the target FPGA.
The results will inform the specification of an eFPGA fabric to be integrated as part of a test chip.

Stats

An 8-bit ADC with 15 MS/s sampling rate can maintain high neutron efficiency at low gamma leakage.
BDT models with max depth of 3 and 50 boosting rounds use less than 3% flip-flops and 25% LUTs on the target Artix 7 FPGA.
fcNN models with a single 16-node hidden layer, 8-bit weights/biases, 12-bit activations, and 30% sparsity use less than 3% flip-flops and 10% LUTs on the target FPGA.

Quotes

"We were able to show that an ADC with only 8-bits resolution and Nyquist sampling rates of approximately 15 MS/s can still allow us to maintain excellent neutron efficiency at very low gamma leakage values."
"For BDT models, a max depth of 3 and 50 boosting rounds were found to be a good balance between performance and resource usage."
"For fcNN models, a single hidden layer of size 16 with 8-bit weights/biases and 12-bit activations, along with 30% sparsity, achieved high neutron efficiency."

Key Insights Distilled From

Investigating Resource-efficient Neutron/Gamma Classification ML Models Targeting eFPGAs

by Jyothisraj J... at arxiv.org 04-24-2024

https://arxiv.org/pdf/2404.14436.pdf

Deeper Inquiries

How can the input feature extraction be further optimized to reduce resource usage while maintaining classification performance

In order to optimize input feature extraction to reduce resource usage while maintaining classification performance, several strategies can be implemented. One approach is to carefully select and engineer the input features to capture the most relevant information for the classification task. This involves conducting a thorough feature selection process to identify the most discriminative features while discarding redundant or irrelevant ones. By reducing the number of input features, the overall resource usage can be minimized without compromising classification performance.
Another optimization technique is to implement feature scaling and normalization to ensure that all input features are on a similar scale. This can help improve the convergence of the machine learning models during training and enhance their overall performance. Additionally, quantization of input features can be optimized by selecting the appropriate bit-resolution that balances resource usage with classification accuracy. By carefully tuning the quantization levels based on the specific requirements of the application, resource-efficient models can be developed.
Furthermore, exploring techniques such as dimensionality reduction through methods like Principal Component Analysis (PCA) or autoencoders can help reduce the complexity of the input feature space while preserving important information. By transforming the input features into a lower-dimensional space, the computational and memory requirements of the models can be reduced, leading to more efficient implementations.

What are the potential trade-offs between BDT and fcNN models in terms of latency, throughput, and power consumption for this application

The potential trade-offs between BDT and fcNN models in terms of latency, throughput, and power consumption for this application depend on the specific requirements and constraints of the deployment environment.
Boosted Decision Trees (BDT) are known for their interpretability, simplicity, and efficiency in handling high-dimensional data. They are well-suited for applications where transparency in decision-making is crucial. BDT models typically have lower latency and higher throughput compared to fully connected Neural Networks (fcNN), making them suitable for real-time applications that require quick decision-making. However, BDT models may require more memory and storage due to the nature of decision trees, which can impact power consumption.
On the other hand, fcNN models are powerful in capturing complex patterns in data and can offer higher accuracy in classification tasks. They are capable of learning intricate relationships between input features, making them suitable for applications with intricate data structures. However, fcNN models often require more computational resources, leading to higher latency and potentially lower throughput compared to BDT models. This increased computational demand can also result in higher power consumption.
The choice between BDT and fcNN models should be based on a careful consideration of the specific requirements of the application, including the need for interpretability, speed of inference, accuracy, and resource constraints. By evaluating these factors, the most suitable model can be selected to optimize performance while balancing latency, throughput, and power consumption.

How can the techniques demonstrated in this work be extended to other radiation detection applications beyond neutron/gamma classification

The techniques demonstrated in this work can be extended to other radiation detection applications beyond neutron/gamma classification by adapting the ML models and input feature extraction methods to suit the specific characteristics of the new application.
For instance, in applications such as radioactive isotope identification or radiation source localization, the ML models can be trained on datasets containing relevant features specific to these tasks. Input feature extraction can be optimized by considering the unique signatures of different radiation sources and designing features that capture these distinctions effectively. By tailoring the ML models to the new application requirements, resource-efficient implementations can be achieved.
Furthermore, the methodology of exploring parameter spaces for eFPGA implementations can be applied to various radiation detection scenarios. By conducting thorough studies on the impact of different hyperparameters, input feature quantization levels, and model architectures, researchers can optimize the performance of ML models while minimizing resource usage. This approach can be valuable in developing efficient and accurate radiation detection systems for a wide range of applications in nuclear physics, medical imaging, environmental monitoring, and security screening.