What is it about?
Embedded AI systems need to learn new tasks over time, just as humans do. However, these systems have very limited memory and processing power. When they learn new information, they often forget what they learned earlier, a problem known as catastrophic forgetting. Existing solutions replay old data to refresh the system’s memory, but these methods require long processing times and repeated compression-decompression steps. As a result, they become slow and power-hungry, making them impractical for resource-constrained embedded AI systems. Our work introduces Replay4NCL, a new methodology that enables continual learning in neuromorphic (brain-inspired) systems much more efficiently. The key idea is to replay compact latent memories from earlier training with lower processing timesteps of the spiking neural network. However, using fewer steps produces less spikes, which can reduce accuracy. To solve this, Replay4NCL adjusts neural threshold levels and learning rates so the network can still learn effectively even with reduced activity. In our experiments, Replay4NCL preserved previous knowledge more accurately than the state-of-the-art, while achieving 4.88x faster processing, 20% memory savings, and 36% lower energy consumption. This makes it a practical solution for enabling continual learning on the next generation of low-power edge AI devices.
Featured Image
Photo by Growtika on Unsplash
Why is it important?
This work is important because AI is moving into everyday devices like smartphones, fitness trackers, mobile robots, and smart home sensors. These devices need to keep learning as things around them change, but they run on batteries and have limited memory and processing power. Existing continual learning methods for neuromorphic models are slow or power-hungry to work on these small devices. Replay4NCL provides a timely solution by preserving past knowledge while learning new tasks with far fewer timesteps, drastically lowering processing and energy demands. Thus, it directly supports the next generation of adaptive, low-power, embedded AI systems. This makes the work relevant to a wide audience across neuromorphic computing, embedded systems, robotics, and edge AI. Its unique combination of efficiency, accuracy, and biological inspiration resonates with readers seeking practical methods for making continual learning feasible in real-world devices.
Perspectives
From my perspective, this work began with a challenge I encountered while studying neuromorphic continual learning for embedded AI. These systems need to learn new information, yet they operate under extremely limited energy, memory, and latency budgets. Existing replay-based approaches deliver good accuracy, but their long processing times and heavy data handling make them difficult to use on small, battery-powered hardware. I wanted to find a way for these systems to keep learning without forgetting, while still staying fast and energy efficient. Developing Replay4NCL showed me that this is possible when we use fewer timesteps, store only compact memories, and carefully adjust the network’s internal parameters. This makes continual learning much more practical for tiny edge devices. I see this work as a step toward bridging the gap between neuromorphic research and real-world deployment—helping future AI systems become both adaptive and efficient simultaneously.
Mishal Fatima Minhas
United Arab Emirates University
Read the Original
This page is a summary of: Replay4NCL: An Efficient Memory Replay-based Methodology for Neuromorphic Continual Learning in Embedded AI Systems, June 2025, Institute of Electrical & Electronics Engineers (IEEE),
DOI: 10.1109/dac63849.2025.11132839.
You can read the full text:
Contributors
The following have contributed to this page
