Runtime-Robust Edge Inference System with Masking-Based Partial Update on Dynamic Reconfigurable FPGA.

Area of Science:

• Edge Computing
• Artificial Intelligence
• Hardware Acceleration

Background:

• Edge inference systems require real-time performance amidst environmental changes like sensor noise and new object classes.
• Static, offline-trained models on edge devices degrade accuracy due to input distribution drift.

Purpose of the Study:

• To propose a runtime-robust edge inference framework enabling continuous adaptation without execution interruption.
• To leverage server-side FPGA acceleration for efficient, dynamic model updates on edge devices.

Main Methods:

• Memory partitioning into active and adaptive regions on the edge device.
• Server-side FPGA performing layer-wise importance analysis and partial retraining.
• Dynamic Partial Reconfiguration (DPR) for adaptive mask generation and minimized reconfiguration delay.

Main Results:

• Adaptation latency reduced by up to 1.3x compared to GPU full retraining.
• Communication costs reduced to 28% of full model transmission.
• Demonstrated real-time adaptability, low latency, and communication efficiency.

Conclusions:

• The proposed framework effectively combines masking-based selective updates with FPGA DPR acceleration.
• Achieves robust, real-time learning and adaptation in cloud-edge collaborative environments.
• Offers a viable solution for dynamic edge inference challenges.