Low-rank adaptation for edge AI

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision
• Edge Computing

Background:

• Edge artificial intelligence (AI) offers transformative applications but faces challenges in updating models on resource-limited devices.
• Updating models on edge devices is hampered by constraints in computation and communication capabilities.
• Existing methods for model updates often require substantial computational and communication resources, making them unsuitable for edge environments.

Purpose of the Study:

• To introduce a novel method, Low-Rank Adaptation for Edge AI (LoRAE), designed to address the challenges of updating AI models on edge devices.
• To significantly reduce the number of trainable parameters required for model updates, thereby minimizing computational and communication overhead.
• To evaluate the effectiveness of LoRAE in maintaining or improving model accuracy across diverse AI tasks on edge devices.

Main Methods:

• LoRAE utilizes low-rank decomposition of convolutional neural network (CNN) weight matrices to decrease the number of updated parameters.
• The method reduces the trainable parameters to approximately 4% compared to traditional full-parameter updates.
• Experiments were conducted on image classification, object detection, and image segmentation tasks using the YOLOv8x model.

Main Results:

• LoRAE achieved substantial parameter reductions: 86.1% for image classification, 98.6% for object detection, and 94.1% for image segmentation.
• These significant reductions in trainable parameters were achieved without compromising, and in some cases enhancing, model accuracy.
• The method effectively mitigates computational and communication challenges associated with updating AI models on edge devices.

Conclusions:

• LoRAE presents a highly efficient and precise solution for updating AI models in resource-constrained edge environments.
• The method demonstrates the feasibility of deploying advanced AI capabilities on edge devices through optimized model updates.
• LoRAE holds significant potential for advancing the practical implementation of edge AI across various applications.