Highly Parallel and Ultra-Low-Power Probabilistic Reasoning with Programmable Gaussian-like Memory Transistors
발표자
()
초록
내용
Probabilistic inference in data-driven models has garnered critical interest for its promising capability to predict outputs, reducing risks arising from overconfidence in learning and sensing. However, implementing complex statistical computations in hardware with minimal devices is still challenging. Here, we propose a Gaussian memory transistor(GMT), with heterojunction of p, n-type semiconductors to induce a Gaussian-like response in the current-voltage relation. A separate floating-gate is inserted into an anti-ambipolar heterojunction transistor, which allows the control of the Gaussian parameter, mean(μ) and standard deviation(σ), enables evaluation of complex distribution functions. GMT device also shows excellent retention performance, cyclic endurance, and mechanical flexibility. The GMT simplifies circuit design for probabilistic inference and achieves higher parallelism. We demonstrate its application in localization and obstacle avoidance using Gaussian curves from GMT.