Microsoft Research Asia
A basic research institution established by Microsoft in the Asia-Pacific region, focused on frontier research in computer science. It has cultivated numerous AI talents in fields such as natural language processing, computer vision, and code intelligence.
Timeline (last 90 days)
After completing his PhD in 2016, Tang Duyu entered Microsoft Research Asia, where he led related research such as CodeBERT and CodeGPT.
Microsoft Research Asia's MInference framework matches the most suitable sparse computation method to different attention heads, achieving up to 10x speedup in the prefill stage of million-word long contexts without reducing retrieval accuracy.
Deformable convolution was first proposed by Dai Jifeng et al. from Microsoft Research Asia at the ICCV 2017 conference (DCNv1), with an improved version DCNv2 released in 2019.
Deformable convolution was proposed by Microsoft Research Asia in 2017 and adaptively focuses on irregularly shaped target regions by learning offsets for each sampling position of the convolution kernel.
All Facts (7)
Qlib是微软亚洲研究院开源的AI量化投资平台，提供从数据处理、因子挖掘、模型训练到回测交易的全流程工具链
90%Unverified微软亚洲研究院做过一个叫SynthLM的框架，专门研究怎么用预训练语料大规模生成合成数据
85%Unverified微软亚洲研究院的SynthLM研究结论是合成数据具有可预测性，可以通过规模法则合理选择模型规模和合成数据量来最大化性能提升
80%UnverifiedAfter completing his PhD in 2016, Tang Duyu entered Microsoft Research Asia, where he led related research such as CodeBERT and CodeGPT.
50%UnverifiedMicrosoft Research Asia's MInference framework matches the most suitable sparse computation method to different attention heads, achieving up to 10x speedup in the prefill stage of million-word long contexts without reducing retrieval accuracy.
50%UnverifiedDeformable convolution was first proposed by Dai Jifeng et al. from Microsoft Research Asia at the ICCV 2017 conference (DCNv1), with an improved version DCNv2 released in 2019.
50%UnverifiedDeformable convolution was proposed by Microsoft Research Asia in 2017 and adaptively focuses on irregularly shaped target regions by learning offsets for each sampling position of the convolution kernel.
50%