New Breakthrough in Embedded AI: How the Replay4NCL Engine Overcomes the Challenges of Continuous Learning?
-Click the blue text to follow us-
Introduction
In the era of booming artificial intelligence, embedded AI systems are gradually becoming the “smart brains” of various intelligent devices, from mobile robots navigating complex environments to drones soaring in the sky, they are ubiquitous. However, as application scenarios become increasingly complex and variable, embedded AI systems face a tricky problem – how to continuously learn in dynamic environments while avoiding forgetting previously learned knowledge, all while ensuring efficient resource utilization. Just as everyone is racking their brains over this dilemma, researchers from the University of the UAE, New York University Abu Dhabi, and the National University of Sciences and Technology in Pakistan have brought good news. Their jointly developed Replay4NCL technology provides an innovative solution to the continuous learning dilemma of embedded AI systems, and this achievement is set to be showcased at the 62nd Design Automation Conference (DAC) in San Francisco in June 2025, generating much anticipation.
1.Temporal Optimization: Breaking Traditional Constraints, Balancing Accuracy and Latency
A major core innovation of Replay4NCL lies in the optimization of timing. In spiking neural networks, timing acts like the “metronome” of the system, determining the rhythm at which neurons process information.
Traditional SNN models often use longer timing, which can ensure higher accuracy but also increases processing latency, posing a significant obstacle for embedded AI systems that require real-time responses.
Researchers found through extensive experiments that reducing the timing from the traditional 100 to 40, while slightly decreasing accuracy, still remains within an acceptable range and significantly reduces processing time. This finding provides a key basis for timing optimization. Moreover, Replay4NCL introduces a data compression-decompression mechanism, compressing potential data during storage and decompressing it during use, significantly reducing memory usage of potential data without losing information, akin to putting data in a “compression suit,” saving a lot of space.
2.Parameter Adjustment: Cleverly Compensating for Information Loss, Enhancing Learning Ability
While reducing timing decreases latency and memory usage, it raises new issues. The reduced number of spikes received by neurons is akin to a person receiving less information, which may make it difficult for the membrane potential to reach the threshold potential, thereby affecting network performance.
To address this issue, Replay4NCL has introduced a parameter adjustment module.
Researchers have lowered the threshold potential Vthr value, making it easier for neurons to fire spikes. Even with a limited number of spikes, they can maintain a spiking activity similar to the original pre-trained model. At the same time, lowering the learning rate slows down the network’s learning speed, allowing the network to update weights more cautiously during the training phase, especially when the number of spikes is low, thus enhancing the retention of old knowledge while also improving the ability to learn new knowledge, akin to adding a “stabilizer” to the network’s learning process.
3.Dynamic Training Strategy: Organically Integrating Innovations to Create an Efficient Learning Mechanism
The dynamic training strategy of Replay4NCL is another highlight, cleverly integrating timing optimization, parameter adjustment, and potential replay data insertion strategies into an efficient training mechanism.
During the pre-training phase, the SNN model first learns all pre-training tasks. Before entering the continuous learning training phase, the model generates LR data activation and divides the network into frozen layers and learning layers based on selected layers. The frozen layers are responsible for transmitting input spikes, while the learning layers are updated when learning new tasks. During the continuous learning training process, the network dynamically adjusts the threshold potential and learning rate, allowing the network to efficiently update weights and learn even with fewer spikes, akin to creating an intelligent “learning engine” for the network that automatically adjusts learning strategies based on different situations.
4.Experimental Validation of Strength: Surpassing in Multiple Aspects, Leading a New Direction for Embedded AI
To validate the performance of Replay4NCL, researchers conducted comprehensive tests in scenarios such as Spiking Heidelberg Digits and Class-Incremental Learning, focusing on key metrics such as accuracy, processing latency, memory usage, and energy consumption.
The experimental results were surprising. Replay4NCL excelled in retaining old knowledge, achieving a Top-1 accuracy of 90.43%, surpassing the existing advanced method SpikingLR by 4.21 percentage points, while maintaining comparable accuracy when learning new tasks. In terms of processing latency, with a timing setting of 40, the latency was reduced by 4.88 times compared to SpikingLR with a timing of 100, greatly enhancing the system’s real-time response capability. In terms of memory usage, thanks to the data compression-decompression mechanism, the memory usage of potential data was reduced by 20%, effectively alleviating the system’s storage burden. Regarding energy consumption, it was reduced by 36.43% compared to SpikingLR, demonstrating significant energy-saving effects, which is crucial for battery-powered embedded devices.
Conclusion:Opening a New Chapter in Embedded AI
The emergence of Replay4NCL brings new hope for continuous learning of embedded AI systems in dynamic environments. Through a series of innovative designs, it successfully addresses the high latency, large energy consumption, and high memory usage issues faced by traditional methods, achieving comprehensive improvements in accuracy, performance, and resource utilization. With the demonstration and promotion of this technology at the 62nd Design Automation Conference, it is believed that it will trigger a new revolution in the field of embedded AI, driving numerous application scenarios such as mobile robots, autonomous driving, and drones towards a more intelligent and efficient direction. Looking forward to the future, Replay4NCL can continuously optimize and improve, injecting more momentum into the development of artificial intelligence, allowing intelligent devices to truly achieve “smart evolution.”
Scan to add “AI Navigator”
Join the community:
“AI Ecological Alliance”
“AI Learning Co-Creation Camp”
“DeepSlash Wealth Community”