Electrostatic Gaussian Transistor for Real-Time Probabilistic Inference

Area of Science:

• Semiconductor device physics
• Probabilistic computing hardware
• Analog circuit design

Background:

• Gaussian distribution functions are crucial for probabilistic computing but challenging to implement in hardware.
• Existing anti-ambipolar transistors suffer from asymmetries, hindering accurate Gaussian curve generation.

Purpose of the Study:

• To report a new single-channel split-gate Gaussian-mirroring transistor (SC-GMT) for faithful hardware implementation of Gaussian functions.
• To demonstrate the device's tunable characteristics and practical applications in probabilistic models.

Main Methods:

• Fabrication of a single-material, single-channel split-gate transistor.
• Utilizing reversal voltage biasing and independent split-gate control for modulation.
• Integration with digital-to-analog control and current sensing for system demonstration.

Main Results:

• The SC-GMT generates symmetric, tunable Gaussian-shaped transfer curves with high fidelity (>99.99% R²).
• A hardware Gaussian Naive Bayes classifier using SC-GMTs achieved 82% accuracy in deepfake voice detection.
• The transistor exhibits quadratic-order analog multiplication capabilities.

Conclusions:

• The SC-GMT offers a robust solution for hardware implementation of Gaussian functions, overcoming limitations of previous approaches.
• This technology enables efficient analog computation for advanced probabilistic models and machine learning architectures.
• The device shows promise for applications in signal processing, pattern recognition, and secure communication.