Memristor-based feature learning for pattern classification.

Area of Science:

• Neuromorphic engineering
• Materials science
• Computer science

Background:

• Deep learning models, inspired by biology, are computationally complex and energy-intensive.
• Existing hardware for deep learning often has a disparity with biological systems, leading to inefficiencies.
• High energy consumption poses a sustainability challenge for the growth of deep learning.

Purpose of the Study:

• To develop a feature learning technique that minimizes the disparity between deep models and hardware.
• To propose a novel approach for implementing feature learning directly using semiconductor physics.
• To reduce the computational complexity and energy consumption of intelligent models.

Main Methods:

• A feature learning technique based on memristor drift-diffusion kinetics was developed.
• Leveraged the dynamic response of a single memristor for feature learning.
• Experimentally implemented the proposed network on 180nm memristor chips for pattern classification.

Main Results:

• Model parameters and computational operations were reduced by 2 and 4 orders of magnitude, respectively, compared to deep models.
• The memristor kinetics-based hardware significantly reduced energy and area consumption compared to memristor-based deep learning hardware.
• Demonstrated effective performance on various dimensional pattern classification tasks.

Conclusions:

• Innovations in hardware physics offer a promising solution for balancing model complexity and performance in intelligent systems.
• Memristor drift-diffusion kinetics provide an efficient and sustainable approach to feature learning.
• Direct implementation of feature learning using semiconductor physics minimizes hardware-model disparity.