Hardware Demonstration of SRDP Neuromorphic Computing with Online Unsupervised Learning Based on Memristor Synapses

Area of Science:

• Neuromorphic Engineering
• Materials Science
• Computer Science

Background:

• Neuromorphic computing offers high speed and energy efficiency for cognitive tasks.
• Memristors are key electronic synapse candidates for neuromorphic systems due to their properties.
• Existing memristor systems have limitations in online unsupervised learning capacity.

Purpose of the Study:

• To construct and demonstrate a memristor-based hardware system for online unsupervised learning.
• To implement the spike-rate-dependent plasticity (SRDP) learning rule at the network level.
• To achieve high recognition accuracy for complex tasks using neuromorphic hardware.

Main Methods:

• Constructed a neuromorphic system with memristor arrays as synapses and CMOS circuits as neurons.
• Implemented online unsupervised learning using the SRDP algorithm.
• Trained and tested the network on the MNIST handwritten digit dataset.

Main Results:

• Achieved online unsupervised learning and weight updates for 10 MNIST handwritten digits.
• Demonstrated recognition accuracy exceeding 90% despite 20% device variation.
• Successfully realized SRDP networks on memristor-based hardware.

Conclusions:

• The developed memristor-based system enables efficient online unsupervised learning.
• This work advances the development of large-scale neuromorphic networks for complex cognitive tasks.
• Memristor hardware shows significant potential for future AI applications.