A Fully Spiking Hybrid Neural Network for Energy-Efficient Object Detection.

Area of Science:

• Artificial Intelligence
• Computer Vision
• Neuromorphic Engineering

Background:

• Traditional Deep Neural Networks (DNNs) face limitations in energy efficiency and robustness for object detection on resource-constrained devices.
• Spiking Neural Networks (SNNs) offer potential for lower power consumption but often struggle with complex tasks and training.
• Hybrid approaches combining SNNs and DNNs are emerging to leverage the strengths of both paradigms.

Purpose of the Study:

• To introduce a novel Fully Spiking Hybrid Neural Network (FSHNN) for energy-efficient and robust object detection.
• To evaluate the performance of FSHNN against conventional DNN-based object detectors, particularly in challenging conditions.
• To assess the uncertainty estimation capabilities of the proposed model.

Main Methods:

• The proposed FSHNN utilizes a Spiking Convolutional Neural Network architecture with leaky-integrate-fire neuron models.
• A hybrid learning strategy combining unsupervised Spike Time-Dependent Plasticity (STDP) and back-propagation (STBP) was employed.
• Monte Carlo Dropout was integrated to quantify prediction uncertainty.

Main Results:

• FSHNN demonstrated superior accuracy compared to DNN-based object detectors.
• The network achieved significantly higher energy efficiency.
• FSHNN exhibited enhanced robustness against noisy input data and limited labeled training data.
• Lower uncertainty error was observed in FSHNN compared to DNNs.

Conclusions:

• The Fully Spiking Hybrid Neural Network (FSHNN) presents a viable solution for energy-efficient and robust object detection.
• FSHNN offers a compelling alternative to traditional DNNs for deployment on edge devices with limited computational resources.
• The model's ability to handle noisy data and uncertainty makes it suitable for real-world applications.