The gate injection-based field-effect synapse transistor with linear conductance update for online training

Area of Science:

• * Materials Science and Engineering
• * Computer Engineering
• * Artificial Intelligence

Background:

• * Neuromorphic computing offers an alternative to traditional von Neumann architecture, necessitating efficient synapse devices for integrated data storage and computation.
• * Three-terminal synapse devices are promising for neuromorphic applications due to stability and controllability, but face challenges like non-linear weight updates and limited dynamic range.
• * Existing devices often lack compatibility with conventional CMOS systems, hindering large-scale integration in crossbar arrays.

Purpose of the Study:

• * To propose and investigate a CMOS-compatible gate injection-based field-effect transistor (FET) for neuromorphic computing.
• * To enhance the linearity of conductance updates in synapse devices by employing thermionic emission.
• * To explore the impact of interfacial layers on the conduction mechanism and device linearity.

Main Methods:

• * Fabrication of a gate injection-based FET incorporating an interfacial layer within the gate stack.
• * Investigation of the device's conduction mechanism by analyzing gate current measurements across varying temperatures.
• * Evaluation of synaptic characteristics and performance in artificial neural network simulations, specifically on the MNIST dataset.

Main Results:

• * The proposed FET device, utilizing thermionic emission, demonstrated significantly improved linearity in conductance updates.
• * The study successfully correlated the conduction mechanism with device linearity by manipulating the interfacial layer.
• * The device achieved superior synaptic characteristics, resulting in a high accuracy of 93.17% in MNIST dataset simulations.

Conclusions:

• * The developed thermionic emission-based FET offers a viable solution for overcoming linearity and compatibility issues in neuromorphic synapse devices.
• * The findings highlight the potential of gate injection-based FETs for creating high-performance, large-scale neuromorphic computing systems.
• * This research contributes to advancing neuromorphic hardware by demonstrating a pathway to more efficient and effective artificial neural network simulations.