Enhancing LiAlO synaptic performance by reducing the Schottky barrier height for deep neural network applications

Area of Science:

• Materials Science
• Neuroscience
• Computer Science

Background:

• Shallow neural networks show promise, but deep neural networks (DNNs) require highly precise and linear synaptic analog weight tuning.
• Existing memristor synapse technologies face challenges in meeting these stringent requirements for realistic DNN applications.
• Action recognition tasks in DNNs demand robust and efficient synaptic functionalities.

Purpose of the Study:

• To fabricate and optimize a Lithium Aluminum Oxide (LiAlOX) memristor synapse for high-precision and linear analog weight tuning.
• To investigate the impact of device optimization on the performance of deep neural networks for action recognition.
• To reduce the nonlinearity factor and enhance the adjustability of synaptic conductance states.

Main Methods:

• Fabrication and optimization of LiAlOX memristor devices.
• Tuning initial conductance states to achieve multiple, continuously adjustable levels.
• Characterization of device performance using current-voltage (I-V) curves to estimate Schottky barrier height (SBH).
• Implementation of the optimized memristor synapses in a deep neural network for action recognition.

Main Results:

• Achieved 120 levels of continuously adjustable conductance states.
• Substantially reduced the nonlinearity factor from 8.96 to 0.83.
• Attributed performance enhancements to a reduced Schottky barrier height (SBH) between the filament tip and electrode.
• Increased action recognition accuracy from 15.1% to 91.4% on the Weizmann video dataset using the optimized devices.

Conclusions:

• The optimized LiAlOX memristor synapse effectively meets the stringent requirements for deep neural networks.
• Device optimization, particularly the reduction of SBH, is crucial for achieving high precision and linearity in synaptic weight tuning.
• The developed memristor synapse technology shows significant potential for advancing realistic artificial intelligence applications, such as action recognition.