Semantic memory-based dynamic neural network using memristive ternary CIM and CAM for 2D and 3D vision

Area of Science:

• Neuroscience
• Computer Science
• Materials Science

Background:

• The human brain excels at dynamic, associative processing, integrating memory and computation.
• Current Artificial Intelligence (AI) models are static, lacking associative capabilities and relying on separated memory and processing units.
• This creates a significant gap between biological intelligence and artificial systems.

Purpose of the Study:

• To propose a hardware-software co-design for a dynamic neural network that mimics the brain's associative capabilities.
• To leverage memristor technology for integrated memory and processing, enabling efficient data association with past experiences.
• To reduce the computational and energy overhead of AI models.

Main Methods:

• Developed a semantic memory-based dynamic neural network architecture utilizing memristors.
• Implemented the network and semantic memory using noise-robust ternary memristor-based computing-in-memory (CIM) and content-addressable memory (CAM) circuits.
• Validated the co-design on ResNet and PointNet++ models for image and 3D point classification using MNIST and ModelNet datasets.

Main Results:

• Achieved classification accuracy comparable to traditional software-based AI models.
• Demonstrated a significant reduction in computational budget: 48.1% for image classification and 15.9% for 3D point classification.
• Reported substantial energy consumption savings: 77.6% for image classification and 93.3% for 3D point classification.

Conclusions:

• The proposed memristor-based hardware-software co-design effectively replicates the brain's associative memory and dynamic processing.
• This approach offers a pathway to more efficient and brain-like artificial intelligence systems.
• The technology shows promise for reducing the computational and energy demands of complex AI tasks.