How Microsoft Research is advancing the state of the art in AI
For 30 years, Microsoft Research (MSR) has been at the forefront of advancing the foundations of computing and translating new scientific understanding into innovative technologies to create value for Microsoft customers and broad benefit to society. MSR has been a pioneer, delivering cutting-edge AI breakthroughs in vision, speech, language, decision making and machine learning, including conversational speech recognition, machine translation, image captioning, and common-sense question answering, and we translate these breakthroughs into products to help Microsoft customers. We’re not inventing for the sake of inventing, we are focused on how the AI systems of today can help people solve real-world challenges.
More recent efforts have focused on developing large-scale models that can process information in increasingly sophisticated ways while also becoming more natural and intuitive to use. Advances in deep learning, coupled with internet-scale datasets and Microsoft Azure’s increasingly powerful AI supercomputing resources, have made it possible to create AI models that perform a broad range of tasks across many different applications.
Microsoft researchers have been working on these problems for years, developing expertise in areas like parallel computation that allows people to more quickly train machine learning models at unprecedented scale. That’s led to innovations like DeepSpeed, an open-source, deep learning optimization library for distributed training that was developed by Microsoft Research and now is used by the broader computing community and ONNX Runtime, which gives high-performance inference support for large Transformer-based models, helping to optimize cost and latency.
We also know we can’t do this work alone, which is why we work across disciplines and geographies; innovate with universities, institutions and companies like OpenAI, Meta, AMD and Hugging Face, an AI community; and contribute to the open-source community to push the industry forward.
Our world-class research community, combined with advancements in deep learning, massive training data, powerful Azure computing infrastructure, dedication to responsibility and partnerships with industry leaders drive our AI evolution.
Research to product
- Microsoft researchers in 2018 were the first to reach a significant milestone in translation, translating news articles from Chinese to English on a commonly used test set. As soon as the team achieved that historic research milestone, they began adapting the model to work in Microsoft Translator, an Azure Cognitive Service that translates a wide variety of texts ranging from historical research documents to travel websites and production manuals.
- In June 2019, Microsoft researchers were the first to develop a machine learning model that surpassed previous benchmarks on the General Language Understanding Evaluation (GLUE) benchmark, which measures mastery of nine different language understanding tasks ranging from sentiment analysis to text similarity and question answering. Bing’s experts incorporated core principles from that research into their own machine learning model, which helped to improve answers and caption generation.
- In 2020, Microsoft researchers built an AI system that can generate accurate image captions. The new model was made available to customers via the Azure Cognitive Services Computer Vision offering, part of Azure AI, enabling developers to use this capability to improve accessibility in their own services. It was also incorporated into Seeing AI and rolled out to Microsoft Word and Outlook for Windows and Mac, and PowerPoint for Windows, Mac and web.
- In 2022, Microsoft researchers developed a new class of AI models called Z-code, boosting even more accuracy in Translator, an Azure Cognitive Service. Microsoft’s Z-code models consistently improved translation quality over current production models, according to common industry metrics.
The post AI Research appeared first on Stories.
source https://news.microsoft.com/2023/04/04/ai-research/
0 Comments