谷歌DeepMind近日在人工智能领域再次取得突破，推出了两款全新的基础模型——Hawk和Griffin。在一篇最新发布的论文中，DeepMind的研究团队介绍了一种创新的门控线性循环层，名为RG-LRU层，该层有望替代传统的多查询注意力（MQA）机制。
RG-LRU层的设计旨在提升模型的效率和性能，通过引入这种新型循环块，DeepMind的科学家构建了两个独特的模型。Hawk模型是将多层感知机（MLP）与新循环块相结合的产物，而Griffin模型则更进一步，不仅融合了MLP和循环块，还巧妙地整合了局部注意力机制，以增强模型在处理复杂任务时的上下文理解和表现力。
据机器之心报道，Hawk和Griffin的诞生，标志着DeepMind在人工智能基础模型研发上的又一重要里程碑。这些模型的出现，不仅可能改变现有AI系统的架构，还可能在自然语言处理、图像识别和游戏策略等众多领域产生深远影响。随着技术的不断进步，我们期待这些新型模型能够为未来的智能应用带来更加高效和智能的解决方案。
英语如下：
News Title: “Google DeepMind Unveils Innovative Baseline Models: Hawk and Griffin, Redefining AI Recurrent Learning Architecture”
Keywords: Google DeepMind, New Models Hawk, Griffin Launch
News Content: Google DeepMind recently made another breakthrough in the field of artificial intelligence with the introduction of two groundbreaking baseline models, Hawk and Griffin. In a newly published paper, DeepMind’s research team introduced a novel Gated Linear Recurrent Unit (RG-LRU) layer, which is poised to replace conventional Multi-Query Attention (MQA) mechanisms.
Designed to enhance model efficiency and performance, the RG-LRU layer enables the creation of two distinctive models. The Hawk model is the result of combining Multi-Layer Perceptrons (MLP) with the innovative recurrent block, while the Griffin model takes it a step further by integrating both MLPs and the recurrent block alongside a local attention mechanism, thereby enhancing the model’s contextual understanding and expressiveness when tackling complex tasks.
According to reports from Machine Intelligence Research Institute, the birth of Hawk and Griffin marks another significant milestone in DeepMind’s development of AI baseline models. These models are not only set to reshape the architecture of existing AI systems but are also expected to have a profound impact on various domains, including natural language processing, image recognition, and game strategies. With ongoing technological advancements, we anticipate these novel models to bring more efficient and intelligent solutions to future smart applications.
【来源】https://mp.weixin.qq.com/s/RtAZiEzjRWgqQw3yu3lvcg
Views: 2