Pruning and quantization algorithm with applications in memristor-based convolutional neural network.

Area of Science:

• Neuromorphic Engineering
• Artificial Intelligence
• Materials Science

Background:

• Memristor-based convolutional neural networks (CNNs) mimic the brain's efficiency but face challenges with increasing complexity.
• Larger networks require more memristors, leading to higher power consumption and larger model sizes.
• Existing memristor CNNs struggle to scale for advanced applications.

Purpose of the Study:

• To propose a novel SBT-memristor-based CNN architecture.
• To develop a hybrid optimization method combining pruning and quantization for memristor CNNs.
• To reduce the size and power consumption of memristor CNNs while maintaining performance.

Main Methods:

• Constructed an SBT-memristor-based CNN utilizing the memristor's thresholding properties.
• Designed memristive in-memory computing, activation, and max-pooling units.
• Applied a hybrid optimization technique integrating network pruning and weight quantization.

Main Results:

• The proposed architecture significantly reduced the number of memristors required.
• Achieved lower power consumption and a compressed network model.
• Demonstrated faster recognition speeds and reduced power usage on the MNIST dataset with minimal precision loss.

Conclusions:

• SBT-memristor-based CNNs offer a viable solution for energy-efficient and compact AI systems.
• Hybrid optimization effectively simplifies memristor CNNs and improves weight representation.
• This approach provides a pathway for complex CNN applications in resource-constrained environments.