谷歌DeepMind近日在人工智能研究领域再次取得突破，推出了两款全新的基础模型——Hawk和Griffin。这两款模型的诞生，源于DeepMind研究人员在最新论文中提出的创新性技术——RG-LRU层。RG-LRU层是一种独特的门控线性循环层，它有望替代传统的多查询注意力（MQA）机制，为循环神经网络的性能提升开辟新途径。
Hawk和Griffin模型正是基于RG-LRU层构建的。Hawk模型巧妙地融合了多层感知机（MLP）与RG-LRU循环块，旨在实现更高效的信息处理和学习能力。而Griffin模型则更进一步，不仅结合了MLP和RG-LRU循环块，还引入了局部注意力机制，这使得Griffin在处理复杂任务时能展现出更强的上下文理解和适应性。
这两款新模型的推出，不仅展示了DeepMind在人工智能基础研究上的持续探索，也预示着未来AI系统在处理序列数据和理解复杂结构信息时将有更大的潜力。随着Hawk和Griffin的面世，我们期待看到更多实际应用中AI性能的显著提升，特别是在自然语言处理、语音识别和图像分析等领域。
英语如下：
News Title: “Google DeepMind Unveils Innovative Models Hawk and Griffin: Redefining AI Infrastructure”
Keywords: DeepMind, Hawk, Griffin
News Content: Google’s DeepMind has recently made another breakthrough in the field of artificial intelligence research with the introduction of two groundbreaking foundational models, Hawk and Griffin. These models stem from the innovative technology, the RG-LRU layer, proposed by DeepMind researchers in their latest paper. The RG-LRU layer, a distinctive gated linear recurrent unit, has the potential to replace conventional multi-query attention (MQA) mechanisms, opening new avenues for enhancing the performance of recurrent neural networks.
The Hawk model is built upon the RG-LRU layer, skillfully integrating multi-layer perceptrons (MLP) with RG-LRU recurrence blocks, aiming to achieve more efficient information processing and learning capabilities. On the other hand, the Griffin model goes a step further. Not only does it incorporate MLPs and RG-LRU blocks, but it also introduces a local attention mechanism. This allows Griffin to demonstrate stronger contextual understanding and adaptability when dealing with complex tasks.
The release of these new models not only highlights DeepMind’s ongoing exploration in AI fundamental research but also foreshadows the increased potential of AI systems in handling sequential data and understanding complex structural information. With the advent of Hawk and Griffin, we anticipate significant performance enhancements in AI applications, particularly in natural language processing, speech recognition, and image analysis.
【来源】https://mp.weixin.qq.com/s/RtAZiEzjRWgqQw3yu3lvcg
Views: 1