Reconfigurable Floating Gate Memristors for High-Accuracy Neuromorphic Computing.

Area of Science:

• Materials Science
• Computer Engineering
• Neuroscience

Background:

• Conventional computing architectures face limitations with increasing data volumes.
• Floating gate memristors offer potential for neuromorphic computing due to integrated memory and computation.
• Emulating nonlinear activation functions is crucial for deep neural network hardware implementation.

Purpose of the Study:

• To present a reconfigurable floating gate memristor (FGM) capable of emulating nonlinear activation functions.
• To demonstrate a dual-function hardware unit for neuromorphic systems.
• To evaluate the performance of FGM-based deep neural networks.

Main Methods:

• Fabrication of a MoS2/hBN/graphene heterostructure-based FGM.
• Characterization of device performance, including state retention and tunable conductance.
• Emulation of rectified linear unit (ReLU) and leaky rectified linear unit (Leaky ReLU) using back gate voltages.
• Integration of FGMs into LeNet and AlexNet architectures for classification tasks.

Main Results:

• The FGM exhibited stable conductive states and 66 linearly tunable conductance states.
• Multilevel conductance tunability was achieved using optical pulses.
• Successful emulation of ReLU and Leaky ReLU activation functions was demonstrated.
• FGM-integrated deep neural networks showed improved inference accuracy on benchmark datasets compared to activation-free models.

Conclusions:

• The developed FGM serves as a dual-function unit, acting as both a tunable synaptic weight and a native nonlinear activation function.
• This innovation paves the way for fully hardware-implemented neuromorphic systems.
• The device performance supports its application in advanced artificial intelligence hardware.