Continual Learning on Edge Field-Programmable Gate Array (FPGA)

Opportunity

This technology allows edge AI systems to learn continuously from streaming data, enhancing their performance and reliability in dynamic and unpredictable environments. By using FGPA-based accelerators and optimisation strategies, the technology lowers the computational cost and power consumption, making it suitable for resource-constrained devices, such as smartphones, wearable devices, and micro aerial vehicles. The technology can also be applied to various fields requiring real-time, scalable object classification, such as augmented reality, robotics, and autonomous driving. It is open-source and compatible with widely used frameworks, such as ROS, OpenCV, and TensorRT, which facilitates easy implementation and customisation by the research community.

Technology

This technology develops FPGA-based accelerator architectures and optimisation strategies for machine-learning algorithms that enable continual learning on edge devices. Continual learning is the ability to learn from streaming data without forgetting previous knowledge or requiring offline model training, which is crucial for edge AI systems in dynamic and unpredictable environments.

The key components include:

• Two FPGA accelerators for continual learning, based on different models: Self-Organising Neural Network (SONN) and Streaming Linear Discriminant Analysis (SLDA).
• The SONN model performs unsupervised learning from embedding features extracted from a CNN model by dynamically growing neurons and connections. The SLDA model performs supervised learning from embedding features and labels by updating a linear projection matrix.
• Design optimisation strategies and runtime scheduling techniques that optimise resource usage, latency, and energy consumption of the FPGA accelerators.
• Real-time learning on-device, providing high scalability to learn a large number of classes.

Applications & Advantages

Applications:

• Augmented reality
• Robotics
• Autonomous driving

Advantages:

• Enables edge AI systems to learn continuously from streaming data without forgetting previous knowledge or requiring offline model training.
• Leverages FPGA-based accelerators and optimisation strategies to lower costs and energy usage.
• Supports real-time learning and is scalable to handle numerous classes.
• Open-source and based on widely used frameworks, such as ROS, OpenCV, and TensorRT.