谷歌DeepMind，全球领先的人工智能研究机构，近日在其最新论文中揭示了两款全新的基础模型——Hawk和Griffin。这些模型标志着AI学习架构的又一重大突破，旨在提升机器学习的效率和性能。
论文中，DeepMind的研究团队提出了一种名为RG-LRU的门控线性循环层，这是一种创新的设计，旨在替代传统的多查询注意力（MQA）机制。RG-LRU层的引入旨在优化信息处理，提高模型的学习能力和记忆效果。
基于这一新设计，DeepMind构建了两个独特的模型。Hawk，融合了多层感知机（MLP）与新循环块，展示了在复杂任务中的高效学习能力。而Griffin则更进一步，不仅结合了MLP与循环块，还引入了局部注意力机制，这使得Griffin在处理大规模数据和复杂场景时，能够实现更为精确和细致的理解。
这些新模型的推出，预示着人工智能领域在模型架构和学习策略上的持续革新。谷歌DeepMind的这一研究进展，无疑将为未来的AI应用，如自然语言处理、图像识别等领域，提供更为强大的工具和可能性。
英语如下：
Title: “Google DeepMind Unveils Innovative Models Hawk and Griffin, Redefining AI Infrastructure”
Keywords: DeepMind, Hawk, Griffin
News Content:
Google DeepMind, a leading global AI research organization, recently unveiled two groundbreaking foundational models, Hawk and Griffin, in its latest paper. These models signify a significant advancement in AI learning architectures, aimed at enhancing the efficiency and performance of machine learning.
In the paper, DeepMind’s research team introduces the Recurrent Gated Linear Unit with Reset (RG-LRU), a novel gated linear recurrent layer designed to replace traditional Multi-Query Attention (MQA) mechanisms. The introduction of RG-LRU layers is targeted at optimizing information processing and improving the learning capacity and memory effectiveness of models.
Based on this innovative design, DeepMind has constructed two distinct models. Hawk, integrating Multi-Layer Perceptrons (MLPs) with new recurrent blocks, demonstrates efficient learning capabilities in complex tasks. Going a step further, Griffin combines MLPs and recurrent blocks along with local attention mechanisms, enabling it to achieve more precise and nuanced understanding when dealing with large-scale data and intricate scenarios.
The release of these new models signals a continuous evolution in AI model architecture and learning strategies. Google DeepMind’s research progress is set to provide more powerful tools and possibilities for future AI applications, such as natural language processing and image recognition.
【来源】https://mp.weixin.qq.com/s/RtAZiEzjRWgqQw3yu3lvcg
Views: 1