Memristor bridge synapse-based neural network and its learning.

Area of Science:

• Neuroscience
• Computer Engineering
• Materials Science

Background:

• Analog neural networks face challenges with nonvolatile weight storage.
• Memristor bridge synapses offer a potential solution for nonvolatile memory in analog circuits.
• Spatial non-uniformity and non-ideal responses of memristors complicate hardware implementation.

Purpose of the Study:

• To propose an analog hardware architecture for a memristor bridge synapse-based multilayer neural network.
• To develop an efficient learning scheme for the proposed hardware architecture.
• To address challenges associated with memristor non-uniformity and non-ideal responses.

Main Methods:

• Designed a multilayer neural network architecture utilizing memristor bridge synapses.
• Developed a modified chip-in-the-loop learning scheme involving initial software training and hardware learning of individual neurons.
• Implemented forward calculations on circuit hardware and weight updates assisted by a host computer.
• Eliminated the need for synaptic weight readout during weight updates.

Main Results:

• Demonstrated nonvolatile weight storage capability of the memristor bridge synapse.
• Successfully implemented the proposed learning scheme on hardware.
• Validated the architecture and learning scheme on a three-bit parity network and a car detection network.

Conclusions:

• The proposed analog hardware architecture with memristor bridge synapses effectively enables nonvolatile weight storage.
• The modified learning scheme compensates for memristor non-idealities and facilitates efficient hardware training.
• The approach shows promise for practical analog neural network implementations.