Nature编辑点评这项研究：智源提出的Emu3仅基于预测下一个词元，实现了大规模文本、图像和视频的统一学习，其在生成与感知任务上的性能可与使用专门路线相当，这一成果对构建可扩展、统一的多模态智能系统具有重要意义。

智源的Emu3模型则开辟了一条新路。该模型基于“预测下一个词元”的全新多模态学习框架，将图像、文本和视频统一离散化到同一个表示空间中。研究团队从零开始，在多模态序列混合数据上联合训练一个单一的Transformer架构，证明了仅凭“预测下一个词元”，就能够同时支持高水平的生成能力与理解能力。

而多模态模型主要依赖对比学习、扩散模型等专门路线，自回归路线是否可以作为通用路线统一多模态？一直是未解之谜。智源这项成果表明，只采用自回归路线，就可以统一多模态学习，训练出优秀的原生多模态大模型，对于确立自回归成为生成式人工智能统一路线具有重大意义。

“Our goal is to build agency in the next generation,” said Lax Poojary, CEO and founder of Sparkli. “Children learn by ...

智源Emu3成果登上Nature：基于“预测下一个Token”,智源,token,模态,序列,实验 ...

LONDON, ENGLAND - APRIL 04: Ai-Da Robot, an ultra-realistic humanoid robot artist, paints during a press call at The British Library on April 4, 2022 in London, England. Ai-Da will open her solo ...

Immune Checkpoint Blockade (ICB) has reshaped cancer care and can deliver durable remission in malignancies such as melanoma and non-small cell lung cancer.

Reflecting on the developments of 2024, this year has been transformative for the entire educational landscape. We’ve witnessed how the thoughtful integration of artificial intelligence can elevate ...

If you have engaged with the latest ChatGPT-4 AI model or perhaps the latest Google search engine, you will of already used multimodal artificial intelligence. However just a few years ago such easy ...