CIM-Based Smart Pose Detection Sensors.

Area of Science:

• Computer Engineering
• Artificial Intelligence
• Sensor Technology

Background:

• Digital sensors often use von Neumann architecture, leading to high energy and resource consumption for continuous, complex computations.
• Advanced sensing algorithms, including deep neural networks, exacerbate these resource demands, limiting 24/7 deployment.

Purpose of the Study:

• To design and evaluate a Computing-In-Memory (CIM) based sensing framework for human pose recognition.
• To address the high energy and resource consumption challenges in always-on sensing systems.

Main Methods:

• Developed a CIM-based sensing framework integrating storage and analog processing to minimize data movement.
• Incorporated uncertainty-aware training, specialized activation function design, and CIM error modeling.
• Evaluated framework performance on human pose classification using binary weights.

Main Results:

• Achieved a significant improvement in pose classification accuracy on CIM devices, rising from 33.3% to 91.5% with binary weights.
• Demonstrated near-ideal CIM accuracy (92.1%) and high digital system accuracy (98.7% binary, 99.5% floating weight).
• Reported 30,000 to 50,000 times lower energy consumption per convolution layer compared to digital systems.

Conclusions:

• The CIM-based framework substantially reduces power consumption for sensing applications.
• This approach enables the development of battery-powered, always-on sensors for continuous monitoring.