Continual Learning on Edge is the deployment of continual/incremental learning algorithms on edge devices — enabling models to learn new tasks or adapt to distribution drift without forgetting previous knowledge, all within the tight resource constraints of edge hardware.

Edge Continual Learning Challenges

• Memory: Cannot store large replay buffers — need memory-efficient continual learning methods.
• Compute: Regularization-based methods (EWC, SI) add minimal compute overhead — suitable for edge.
• Storage: Cannot keep full copies of past models — need compact knowledge summaries.
• Methods: Experience replay (tiny buffer), parameter isolation, knowledge distillation, elastic weight consolidation.

Why It Matters

• Process Drift: Semiconductor processes drift over time — edge models must adapt without redeployment.
• New Products: When new products are introduced, edge models must learn new classes without forgetting old ones.
• Autonomous: Edge devices in remote locations must learn continuously without human intervention.

Continual Learning on Edge is never stop learning, never forget — enabling edge devices to continuously adapt while maintaining knowledge of past tasks.