手の動きを感知し記憶を保存する新型神経型デバイスを開発(Tiny device promises new tech with a human touch)

ad

2025-05-13 ロイヤルメルボルン工科大学(RMIT)

RMIT大学の研究チームは、外部コンピュータを必要とせず、人間の脳のように情報を処理・記憶できる小型の「ニューロモルフィック」デバイスを開発しました。このデバイスは、モリブデンジスルフィド(MoS₂)という金属化合物を用いており、手の動きを検出し、視覚情報を即座に処理・保存する能力を持ちます。この技術は、自動運転車や先進的なロボット工学など、次世代の人間とのインタラクションを向上させる応用が期待されています。

<関連情報>

光活性単分子膜MoS2によるスパイキングニューラルネットワークを用いたマシンビジョンアプリケーション Photoactive Monolayer MoS2 for Spiking Neural Networks Enabled Machine Vision Applications

Thiha Aung, Sindhu Priya Giridhar, Irfan H. Abidi, Taimur Ahmed, Akram AI-Hourani, Sumeet Walia
Advanced Materials Technologies  Published: 23 April 2025
DOI:https://doi.org/10.1002/admt.202401677

手の動きを感知し記憶を保存する新型神経型デバイスを開発(Tiny device promises new tech with a human touch)

Abstract

Neuromorphic vision and information processing is a strongly emerging avenue for next-generation computing and sensing technologies. Spiking neural networks (SNNs), which mimic the spiking behavior of biological neurons, offer energy efficiency and biological plausibility. The leaky integrate-and-fire (LIF) model is a fundamental building block of SNNs, where incoming electrical pulses are integrated until threshold is reached, triggering a spike and resetting the membrane potential. While several photoactive materials have been explored for basic neural functionalities, the realization of accurate LIF neuron behavior including replication of membrane potential and its application in neuromorphic vision tasks remain unexplored. This work demonstrates the potential of chemical vapor deposition grown visible-active, atomically thin molybdenum disulfide (MoS2) monolayers in replicating the charging and discharging states of LIF neurons through their photoelectrical response. Gate voltage modulation is employed to swiftly reset the neuron, enabling the replication of LIF dynamics and fast system responsiveness. The key photoresponse parameters of MoS2 are embedded into SNN model that achieves 75% accuracy after 15 epochs for static tasks and 80% accuracy after 60 epochs for dynamic tasks. This demonstrates that the potential of photoactive 2D materials for SNNs, representing a step in the development of real-time, energy efficient neuromorphic computing.

1601コンピュータ工学
ad
ad
Follow
ad
タイトルとURLをコピーしました