Evaluating Continual Learning Techniques in Edge AI for Real-Time Applications
Dr. Farheen Sultana,
Mohd Zabih,
Golla Janardhan
This research investigates the efficacy of various continual learning techniques—Elastic Weight
Consolidation (EWC), Experience Replay, and Knowledge Distillation—within the framework of Edge
AI for real-time applications. Continual learning, crucial for adapting AI models to new data while
preserving previously acquired knowledge, presents unique challenges when deployed on resourceconstrained
edge devices. This study evaluates these techniques based on key performance metrics
including task accuracy, old task accuracy, latency, resource usage, and adaptability. The findings reveal
that Experience Replay excels in maintaining high task accuracy and adaptability, albeit with increased
resource demands. EWC provides a balanced approach with moderate performance and resource usage
but shows slightly lower adaptability. Knowledge Distillation offers an efficient solution with good
performance and minimal computational overhead, making it suitable for edge environments with strict
resource constraints. These insights guide the selection of continual learning methods tailored to the
specific needs of real-time Edge AI applications.
