Gradient Backpropagation based Feature Attribution to Enable Explainable-AI on the Edge

Feature attribution methods can run efficiently on edge devices: The paper demonstrates that explainability algorithms don't require entirely separate hardware—existing inference accelerators can be repurposed with "minimal overhead" to provide real-time explanations, making XAI practical for deployment on resource-constrained edge devices rather than just in data centers.

Multiple explanation approaches can share the same hardware: The researchers developed a flexible FPGA design supporting three different gradient-based feature attribution algorithms, showing that rather than building separate systems for each explanation method, a single configurable accelerator can handle multiple approaches to interpreting AI decisions.

Memory efficiency is critical for on-device explainability: The paper identifies gradient computation memory overhead as a key challenge and proposes optimization strategies to minimize it, highlighting that memory constraints—not just processing power—are a significant barrier to bringing explainability tools to edge devices like smartphones or IoT systems.

Practical implementation validates the approach: The researchers tested their design on real Xilinx FPGAs using standard neural networks trained on CIFAR-10, demonstrating that this isn't purely theoretical work but represents a concrete pathway toward deploying explainable AI systems in real-world edge computing scenarios.