近日，谷歌旗下的AI研究公司DeepMind在学术论文中介绍了其最新研究成果——两个全新的人工智能基础模型「Hawk」和「Griffin」。这两个模型均采用了DeepMind自主研发的RG-LRU层，这一创新性设计使得模型在处理复杂任务时表现更为出色。
RG-LRU层是一种门控线性循环层，它结合了门控机制和线性循环的优点，能够有效提高模型的表征能力和鲁棒性。基于RG-LRU层，DeepMind设计了一种新型的循环块，用以替代现有的多查询注意力（MQA）机制。这一循环块不仅提高了模型处理序列数据的能力，还使其在混合型任务中表现更为优异。
新模型「Hawk」是一款集成了MLP和循环块的混合型模型，它在保持MLP强大表征能力的同时，通过引入循环块提高了处理序列数据的能力。而「Griffin」模型则在「Hawk」的基础上，进一步融合了局部注意力机制，使得模型在处理长序列数据时更为精准。
DeepMind的这一研究成果标志着人工智能领域在模型设计和优化方面的又一次重要突破。据悉，这两个新模型已经在多个任务中展现了优异的性能，有望为自然语言处理、图像识别等领域带来新的突破。
作为人工智能领域的领军企业，谷歌DeepMind的这一成果将进一步推动AI技术的发展，为各行各业带来更多创新应用。未来，我们有理由相信，随着AI技术的不断进步，人工智能将在更多领域发挥重要作用，助力人类社会的发展。
英语如下：
**Title:** DeepMind Unveils New AI Models ‘Hawk’ and ‘Griffin’
**Keywords:** DeepMind, New Models, Hawk, Griffin
**News Content:**
#### Google’s DeepMind Unveils New AI Fundamentals Models ‘Hawk’ and ‘Griffin’
Recently, Google-owned AI research firm DeepMind introduced its latest research findings in academic papers—two brand new artificial intelligence fundamental models named ‘Hawk’ and ‘Griffin.’ Both models adopt the RG-LRU layer, independently developed by DeepMind, which innovatively combines the advantages of gated mechanisms and linear loops, leading to enhanced performance in handling complex tasks.
The RG-LRU layer is a gated linear recurrent unit that merges the benefits of gated structures and linear recurrence, effectively enhancing the model’s representational capabilities and robustness. Based on the RG-LRU layer, DeepMind has designed a new type of recurrent block, replacing the existing multi-query attention (MQA) mechanism. This recurrent block not only improves the model’s ability to process sequential data but also makes it perform better in mixed tasks.
The new model ‘Hawk’ is a hybrid model integrating MLP and recurrent blocks. While maintaining the MLP’s strong representational abilities, it introduces recurrent blocks to enhance the model’s sequential data processing capabilities. The ‘Griffin’ model, built upon ‘Hawk,’ further integrates local attention mechanisms, making the model more precise in processing long sequential data.
This research achievement by DeepMind signifies another significant breakthrough in the field of artificial intelligence in model design and optimization. It is reported that both new models have demonstrated excellent performance in multiple tasks and are expected to bring new breakthroughs to fields such as natural language processing and image recognition.
As a leader in the field of artificial intelligence, Google DeepMind’s achievement will further propel the development of AI technology, bringing more innovative applications to various industries. In the future, it is reasonable to believe that with the continuous advancement of AI technology, artificial intelligence will play a crucial role in more areas, contributing to the development of human society.
【来源】https://mp.weixin.qq.com/s/RtAZiEzjRWgqQw3yu3lvcg
Views: 1