Enhancing memristor multilevel resistance state with linearity potentiation via the feedforward pulse scheme

Area of Science:

• Materials Science
• Computer Science
• Electrical Engineering

Background:

• Artificial intelligence (AI) applications benefit from mapping Artificial Neural Network (ANN) weights to analog memristor resistances for enhanced throughput and energy efficiency.
• Implementing AI on analog memristors faces challenges due to non-linear resistance switching and limited numerical bit precision, impacting ANN model accuracy.

Purpose of the Study:

• To introduce a precise resistance control method for analog memristors.
• To enhance the number of programmable resistance levels for improved AI computation.
• To evaluate the impact of multi-level resistance states on ANN accuracy.

Main Methods:

• A feedforward pulse scheme was developed to precisely control memristor resistance.
• A TiO2-based memristor was used to demonstrate the resistance control method.
• An evaluation framework was established to assess ANN accuracy with varying resistance states.

Main Results:

• The method achieved 512 programmable resistance states on a TiO2-based memristor with a low resistance ratio (1.19).
• Demonstrated 95.5% accuracy on ResNet-34 (over 20 million parameters) via weight transfer using analog memristors.
• Successfully mapped ANN weights to memristor resistances, showcasing potential for AI inference.

Conclusions:

• The developed resistance control method significantly enhances analog memristor precision and programmability for AI.
• Analog memristors show great potential for accurate AI model inference, especially for edge computing applications.
• Future work can focus on increasing resistance states for more complex AI tasks and in-memory computing capabilities.