Pattern classification by memristive crossbar circuits using ex situ and in situ training

Area of Science:

• Materials Science
• Computer Science
• Electrical Engineering

Background:

• Memristors, or memory resistors, offer potential for efficient synaptic weight implementation in artificial neural networks.
• While individual memristor synaptic functions are demonstrated, implementing complete networks remains a challenge.

Purpose of the Study:

• To demonstrate pattern classification using a memristor-based single-layer perceptron network.
• To compare ex situ and in situ training methods for memristive neural networks.

Main Methods:

• A memristive crossbar circuit utilizing titanium dioxide memristors was employed.
• Synaptic weights were realized as memristor conductances.
• Pattern classification was achieved using the perceptron learning rule with both ex situ (pre-calculated weights) and in situ (parallel weight adjustment) training methods.

Main Results:

• Successful pattern classification was demonstrated using the memristor-based perceptron network.
• Both ex situ and in situ training methods yielded satisfactory results, despite memristor variability.
• The study highlights the feasibility of implementing neural networks with memristors.

Conclusions:

• Memristive crossbar circuits can effectively implement artificial neural networks for pattern classification.
• The findings support the development of dense, high-performance neuromorphic computing systems.
• This research advances the practical application of memristors in artificial intelligence hardware.