Random sketch learning for deep neural networks in edge computing.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Engineering

Background:

• Deep neural networks (DNNs) offer significant potential but demand substantial computational resources and memory, hindering deployment on low-cost edge devices.
• Existing lightweight DNN methods face challenges in bridging the resource gap for practical tiny artificial intelligence implementation.
• The need for efficient AI solutions on resource-constrained hardware is critical for broader scientific and industrial adoption.

Purpose of the Study:

• To introduce a novel architecture, Random Sketch Learning (Rosler), for computationally efficient tiny artificial intelligence.
• To develop a universal framework for compressing models while training, enabling direct learning of compact AI models.
• To facilitate computationally efficient on-device learning for AI applications.

Main Methods:

• Developed a compressing-while-training framework named Random Sketch Learning (Rosler).
• Implemented a universal framework that directly learns compact models, optimizing for resource efficiency.
• Validated the approach on diverse models and datasets to assess performance and efficiency gains.

Main Results:

• Achieved substantial memory reduction of approximately 50-90× with 16-bits quantization compared to fully connected DNNs.
• Demonstrated significant acceleration of computation by over 180× on low-cost hardware.
• Reduced energy consumption by approximately 10× on deployed edge devices.

Conclusions:

• Random Sketch Learning (Rosler) offers a viable solution for deploying efficient tiny artificial intelligence on edge devices.
• The developed framework enables significant memory, computational, and energy savings, overcoming previous limitations.
• This method paves the way for widespread adoption of AI in various scientific and industrial applications with resource constraints.