BLADE: ReliaBle Dynamic Hardware-Aware SNN-ANN Boundary SeLection for Event-BAseD Object DEtection

BLADE (ReliaBle Dynamic Hardware-Aware SNN–ANN Boundary SeLection for Event-BAseD Object DEtection) is a framework that introduces reliability-aware boundary selection for dynamic hybrid Spiking Neural Network (SNN) and Artificial Neural Network (ANN) architectures. Unlike previous methods that prioritize only accuracy and energy, BLADE jointly optimizes the SNN–ANN boundary and ANN early-exit configuration based on reliability, detection accuracy, execution time, and energy consumption.

Key capabilities and results include: Dynamic Inference: The architecture uses ANN early exits to adaptively adjust computation based on input complexity, while the SNN backbone processes every input to preserve temporal features. Reliability Optimization: It employs hierarchical statistical fault injection to quantify reliability and identifies dominant failure points, such as specific floating-point exponent bits. Performance Metrics: The deployed hybrid configuration achieves a mean Average Precision (mAP@0.5) of 0.691 with an inference compute energy of 15.82 mJ when the early exit fires. Fault Tolerance: Increasing the SNN portion improves robustness; the fully SNN configuration achieves a reliability retention of 0.965 under aggressive fault conditions, while selective bit-level protection can eliminate catastrophic failures with only ~3% storage overhead.

This methodology enables more dependable deployment of hybrid SNN–ANN systems for safety-critical edge AI applications by allowing trade-offs between dependability and efficiency tailored to specific application requirements.

This arXiv preprint addresses a central tradeoff in event-based object detection: how to exploit the low-power, temporally sparse computation style of Spiking Neural Networks (SNNs) while retaining the strong detection accuracy typically provided by conventional Artificial Neural Networks (ANNs). Hybrid SNN–ANN architectures are attractive because event cameras produce asynchronous, sparse data that can be processed efficiently in early spiking stages, yet downstream detection often benefits from the representational power of ANNs. The paper identifies a key limitation in prior work: many hybrid systems use a static SNN–ANN boundary, where the point at which the network switches from SNN to ANN is fixed in advance and chosen mainly from offline accuracy–energy considerations. Such static designs can be suboptimal because event-rate statistics, scene complexity, and hardware operating conditions vary widely across real-world deployment.

The proposed work, BLADE, introduces a dynamic, hardware-aware boundary selection mechanism for hybrid SNN–ANN event-based detectors. Rather than treating the SNN–ANN split as a fixed architectural choice, BLADE frames it as a runtime decision that can be adjusted according to incoming event data, detection difficulty, and device constraints such as energy, latency, memory, or compute budget. The central insight is that reliable detection does not require a single fixed hybrid topology; instead, the system can allocate more spiking computation when the scene is simple or energy-constrained, and shift more processing to the ANN when accuracy is more critical. This makes the inference pipeline adaptive, resource-aware, and better suited to edge deployment.

The significance of the work lies in making hybrid SNN–ANN detection more practical for low-power, real-time perception systems. Event-based cameras are promising for robotics, autonomous vehicles, and other embedded applications because they can provide high temporal resolution with low power, but only if the downstream network can operate within tight energy and latency budgets. By coupling boundary selection with hardware awareness and reliability considerations, BLADE offers a path toward detectors that maintain robust accuracy while avoiding unnecessary ANN computation. In doing so, it helps bridge neuromorphic event processing and conventional deep detection, moving hybrid SNN–ANN architectures closer to deployable, energy-efficient edge perception.