Transistor-Level Activation Functions via Two-Gate Designs: From Analog Sigmoid and Gaussian Control to Real-Time...

Area of Science:

• Neuromorphic Engineering
• Solid-State Devices
• Artificial Intelligence Hardware

Background:

• Energy-efficient artificial intelligence (AI) hardware relies on tunable analog activation functions.
• Existing solutions often lack precise control over activation function parameters.

Purpose of the Study:

• To introduce novel transistor designs for tunable analog activation functions.
• To demonstrate device-level control over sigmoid and Gaussian function parameters.
• To validate the system-level performance of these transistors in AI hardware.

Main Methods:

• Development of sigmoid-like activation function transistors (SA-transistors) and Gaussian-like activation function transistors (GA-transistors) using a screen gate structure.
• Precise tuning of activation function parameters (slope, saturation, amplitude, standard deviation) via screen gate voltage.
• Integration of transistors into a hardware-based multilayer perceptron (MLP) for system-level validation.

Main Results:

• SA-transistors and GA-transistors demonstrated precise and continuous tunability of analog activation parameters.
• Improved lung MRI classification accuracy from 77% to 84% using SA-transistors.
• Enhanced time-series forecasting R² from 0.82 to 0.93 using GA-transistors.
• Achieved 96.7% accuracy on the IRIS dataset using a hardware-based MLP.

Conclusions:

• Tunable analog activation transistors offer a pathway to hardware-optimized neural computations.
• These devices enable neuromorphic accelerators with reduced circuit complexity and power consumption.
• System-level validation confirms the potential for high classification fidelity without digital post-processing.