Binarized Neural Network with Silicon Nanosheet Synaptic Transistors for Supervised Pattern Classification.

Area of Science:

• Neuromorphic Computing
• Artificial Intelligence Hardware
• Solid State Devices

Background:

• Biological neural networks learn via analog synaptic connections.
• Emerging synaptic devices aim to emulate biological networks for neuromorphic computing.
• Current synaptic device technology faces challenges in on-chip, large-scale artificial neural network implementation due to unreliable analog weight modulation.

Purpose of the Study:

• To demonstrate a binarized neural network (BNN) utilizing gate-all-around silicon nanosheet synaptic transistors.
• To showcase reliable digital-type weight modulation for improved artificial neural network sustainability.
• To validate the feasibility of BNNs for practical applications.

Main Methods:

• Development of a binarized neural network (BNN) architecture.
• Integration of gate-all-around silicon nanosheet synaptic transistors for digital-type weight modulation.
• Application of the BNN to handwritten digit (MNIST), face image (Yale), and custom binary pattern classification tasks using a supervised online training procedure.

Main Results:

• Successful demonstration of a BNN with reliable digital-type weight modulation.
• Proof-of-concept validation across diverse datasets including MNIST, Yale, and a 3x3 binary pattern classification.
• Implementation of a large-scale integrated synaptic transistor network (9x9x2x162 cells).

Conclusions:

• The study consolidates the feasibility of binarized neural networks for artificial intelligence.
• The developed synaptic transistor technology paves the way for reliable, large-scale artificial neural networks.
• Advanced conventional digital device technologies can be leveraged for future neuromorphic computing architectures.