Memristor Based Binary Convolutional Neural Network Architecture With Configurable Neurons.

Area of Science:

• Neuromorphic Engineering
• Computer Science
• Materials Science

Background:

• Memristor-based convolutional neural networks (CNNs) offer low power consumption and high integration for wearable systems.
• Large-scale implementation of memristive devices is hindered by manufacturing limitations and the complexity of high-precision neuron activation functions.

Purpose of the Study:

• To propose a configurable full-binary convolutional neural network (CFB-CNN) architecture to overcome the limitations of current memristor-based CNNs.
• To evaluate the performance of the CFB-CNN architecture under various non-ideal conditions, including device yield and resistance fluctuations.
• To demonstrate the Spiking Neural Network (SNN) compatibility of the memristor-based CFB-CNN.

Main Methods:

• Developed a configurable full-binary convolutional neural network (CFB-CNN) architecture with binary inputs, weights, and neurons.
• Configured neurons in two modes to accommodate non-ideal device behavior.
• Validated the architecture's performance on the MNIST dataset, analyzing the impact of device yield and resistance fluctuations.
• Verified SNN compatibility by encoding pixel values using pulse counts.

Main Results:

• The 2-layer CFB-CNN achieved approximately 98.2% recognition accuracy on the MNIST dataset.
• Under a 64% device yield and ±1 MD neuron configuration, accuracy was 91.28%.
• With 26% resistance variation and 01 MD neuron configuration, accuracy reached 93.43%.
• The memristor-based CFB-CNN demonstrated SNN compatibility.

Conclusions:

• The proposed CFB-CNN architecture effectively addresses the challenges of implementing memristor-based neural networks in practical applications.
• The architecture shows resilience to device imperfections, making it suitable for wearable embedded systems.
• The CFB-CNN's SNN compatibility opens avenues for advanced neuromorphic computing applications.