A linear compensation method for inference accuracy improvement of memristive in-memory computing.

Area of Science:

• Computer Engineering
• Materials Science

Background:

• Memristive computing systems (MCS) offer low power and high parallelism for artificial neural networks (ANNs), presenting an alternative to traditional Von Neumann architectures.
• Non-idealities in peripheral circuits and memristor arrays significantly degrade the performance of practical MCS.

Purpose of the Study:

• To propose and validate a linear compensation method (LCM) for improving MCS performance under non-ideal conditions.
• To investigate and model the output error of MCS considering various non-ideal states.

Main Methods:

• A mathematical model for output error was established based on experimental data, considering the MCS as a whole.
• Physical circuit-level analysis was performed to understand how non-idealities affect output current.
• The proposed LCM compensates output current in real-time using the established mathematical model.

Main Results:

• The LCM effectively compensates for non-idealities in MCS, significantly improving system performance.
• Outstanding performance was demonstrated using the ResNet-34 network architecture, which is sensitive to hardware non-idealities.
• The LCM integrates seamlessly into MCS operations, enhancing deployment on generic ANN hardware.

Conclusions:

• The proposed LCM is a viable solution for mitigating non-idealities in memristive computing systems.
• This method enhances the practical applicability of MCS for ANNs, particularly in hardware implementations.
• The LCM approach facilitates the optimization of memristor-based ANN hardware.