Highly parallel and ultra-low-power probabilistic reasoning with programmable gaussian-like memory transistors.

Area of Science:

• Materials Science and Engineering
• Computer Engineering
• Artificial Intelligence

Background:

• Probabilistic inference in data-driven models is crucial for accurate predictions and confidence estimation, mitigating risks of overconfidence.
• Implementing complex probabilistic computations on resource-constrained devices remains a significant technological challenge.

Purpose of the Study:

• To propose and demonstrate a novel Gaussian-like memory transistor capable of performing probabilistic inference.
• To enable efficient and low-power hardware implementation of complex distribution functions for AI tasks.

Main Methods:

• Fabrication of a heterojunction transistor using p- and n-type semiconductors with a separate floating-gate configuration.
• Achieving a programmable Gaussian-like current-voltage response within a single device.
• Demonstrating the device's performance in localization and obstacle avoidance tasks.

Main Results:

• A single 3-terminal Gaussian-like memory transistor was successfully developed, exhibiting programmable current-voltage characteristics.
• The device demonstrated excellent retention performance (>10000 s) and mechanical flexibility.
• Successful implementation in localization and obstacle avoidance tasks validated the device's utility for probabilistic inference.

Conclusions:

• The proposed Gaussian-like memory transistor offers a promising hardware platform for efficient probabilistic inference computing.
• Its ultralow-power consumption, simplified design, and programmable outputs facilitate complex computations on minimal devices.
• This technology holds potential for advancing AI applications requiring reliable confidence levels and risk mitigation.