谷歌DeepMind近日在一篇学术论文中揭示了其在人工智能领域的最新突破，推出了两个全新的基础模型——Hawk与Griffin。这两款模型是基于研究团队提出的创新性RG-LRU层，这是一种独特的门控线性循环层，旨在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学习效率和性能，它构建了一个全新的循环块，为深度学习架构带来了革命性的变化。基于这一创新，DeepMind构建了两个不同架构的模型。Hawk模型融合了多层感知机（MLP）与RG-LRU循环块，旨在实现更高效的信息处理和学习。而Griffin模型则更进一步，不仅结合了MLP和循环块，还引入了局部注意力机制，这将可能增强模型在处理复杂任务时的精细化理解和响应能力。
这一系列的创新标志着谷歌DeepMind在人工智能基础研究上的持续探索和突破，Hawk和Griffin的发布预示着未来AI模型在处理复杂信息和任务时可能展现出更高的智能水平。这两款模型的出现，无疑将对人工智能领域产生深远影响，为机器学习和自然语言处理的研究提供新的方向。
英语如下：
**News Title:** “Google DeepMind Unveils Innovative Models Hawk and Griffin, Redefining AI Infrastructure”
**Keywords:** DeepMind, Hawk, Griffin
**News Content:**
Google DeepMind recently unveiled groundbreaking advancements in artificial intelligence with the introduction of two novel foundational models, Hawk and Griffin, in a recent academic paper. These models are rooted in the team’s innovative RG-LRU layer, a distinctive gated linear recurrent unit designed to replace conventional multi-query attention (MQA) mechanisms.
The RG-LRU layer has been conceived to enhance learning efficiency and performance, constructing a novel recurrent block that brings a revolutionary shift to deep learning architectures. Based on this innovation, DeepMind has crafted models with distinct architectures. The Hawk model integrates multi-layer perceptrons (MLPs) with RG-LRU recurrent blocks, aiming for more efficient information processing and learning. On the other hand, the Griffin model pushes the boundaries further by combining MLPs, recurrent blocks, and local attention mechanisms, potentially enhancing the model’s ability to understand and respond to complex tasks with greater precision.
These innovations signify Google DeepMind’s ongoing exploration and breakthroughs in fundamental AI research. The introduction of Hawk and Griffin foreshadows a new level of intelligence in AI models when dealing with complex information and tasks. Undoubtedly, these models will have a profound impact on the field of AI, offering new directions for research in machine learning and natural language processing.
【来源】https://mp.weixin.qq.com/s/RtAZiEzjRWgqQw3yu3lvcg
Views: 1