Multimodal transistors as ReLU activation functions in physical neural network classifiers

Area of Science:

• Materials Science
• Computer Science
• Electronics Engineering

Background:

• Artificial neural networks (ANNs) are increasingly integrated into power-efficient thin-film electronics.
• Thin-film technologies demand robust, manufacturable devices with efficient layouts.
• The rectified linear unit (ReLU) is a key activation function in convolutional ANNs (CNNs).

Purpose of the Study:

• To demonstrate that multimodal transistors (MMTs) can emulate the ReLU activation function.
• To assess the impact of MMT transfer characteristics on CNN classification performance.
• To explore the feasibility of using MMTs in practical, power-efficient electronic systems.

Main Methods:

• Simulated and measured transfer characteristics of MMTs were analyzed.
• The linear dependence in MMT saturation was identified as mimicking the ReLU function.
• MATLAB was used to evaluate CNN performance with distorted ReLU functions and MMT proxies.

Main Results:

• MMT transfer characteristics effectively replicate the ReLU activation function.
• High CNN classification accuracy was maintained even with significant variations in MMT parameters.
• The study confirmed that CNNs can utilize MMTs as activation functions due to consistent training and classification behavior.

Conclusions:

• MMTs offer a viable pathway for implementing sophisticated, power-efficient ANNs in thin-film electronics.
• The inherent robustness of MMTs to parameter variations supports their use in real-world applications.
• This research bridges the gap between advanced electronic device physics and artificial intelligence implementation.