韩国研究人员开发双输出突触以实现高效AI

韩国大学的研究人员开发了一种双输出人工突触,以提升多任务AI系统的能量效率,该大学宣布。该设备同时发出电信号和光信号,实现并行处理。测试显示,计算速度提高高达47%,与传统GPU硬件相比,能耗降低最多32倍。

韩国大学KU-KIST融合科学技术的研究生院的研究人员,在王建旭和朴英兰教授的带领下,创建了一种受大脑启发的 artificial synapse。传统的AI芯片是为特定功能设计的,对于多个任务往往需要顺序处理,从而增加功耗。这种新设备通过同时发出电信号和光信号,在单个芯片上允许任务的并行处理。 团队报告在约1000个不同状态下实现了稳定的学习。在测试中,它实现了计算速度提高高达47%,与基于GPU的加速器相比,能耗降低高达32倍。 “这一成就通过同时利用电信号和光信号的人工突触,为多任务AI呈现了一种新的硬件架构,”王教授表示。“它可以进一步扩展到需要复杂决策的高速、低功耗AI系统,例如机器人、医疗保健应用和自动驾驶等领域。” 该研究于周五发表在《Science Advances》上,这是美国科学促进会(American Association for the Advancement of Science)的一本期刊。随着AI对计算能力和电力需求的增加,模仿大脑效率的努力正在获得更多关注。

相关文章

Illustration of glowing whole-brain neural networks coordinating efficiently, representing a University of Notre Dame study on general intelligence.
AI 生成的图像

Study points to whole-brain network coordination as a key feature of general intelligence

由 AI 报道 AI 生成的图像 事实核查

University of Notre Dame researchers report evidence that general intelligence is associated with how efficiently and flexibly brain networks coordinate across the whole connectome, rather than being localized to a single “smart” region. The findings, published in Nature Communications, are based on neuroimaging and cognitive data from 831 Human Connectome Project participants and an additional 145 adults from the INSIGHT Study.

Northwestern University researchers report they have printed flexible “artificial neurons” that generate realistic electrical spike patterns and can trigger responses in living mouse brain tissue. The team says the work, published April 15 in Nature Nanotechnology, could help advance brain-machine interfaces and more energy-efficient, brain-inspired computing.

由 AI 报道

Chinese scientists have drawn inspiration from the Japanese paper-cutting art of kirigami to develop stretchable microelectrode arrays, aiming to overcome limitations in electrode technology such as that used by Neuralink. These arrays were implanted into macaque monkeys, where they flexed with brain tissue to record hundreds of neurons simultaneously. The research was published in the February 5 issue of Nature Electronics.

An Australian company has enabled a chip with human brain cells to play the video game Doom using a simple programming interface. Developed by Cortical Labs, the technology allows for quick training and marks progress toward practical biological computing applications. Experts highlight its potential for handling complex tasks like robotic control.

由 AI 报道

South Korea's science and education ministries discussed on Thursday ways to foster artificial intelligence (AI) talent through programs in elementary and middle schools. The meeting focused on joint efforts to provide AI education across all school stages, including platforms for students to engage directly with data and AI models.

此网站使用 cookie

我们使用 cookie 进行分析以改进我们的网站。阅读我们的 隐私政策 以获取更多信息。
拒绝