They say a picture is worth a thousand words. Sure, but the real trick is realizing a bot that draws pictures is using only a dozen. Throw in an ability to visualize an entire story and one day this bot could be working in the movies: aka.ms/AA5c681 #CVPR2019

Check out my latest article: ConvLab: Multi-Domain End-to-End Dialog System Platform linkedin.com/pulse/convlab-… via LinkedIn

Check out my latest article: The Multi-domain Task Completion track at DSTC8 linkedin.com/pulse/multi-do… via LinkedIn

Congrats to the Dynamics 365 and Microsoft Research AI team for being the first achieve the human performance estimate on the GLUE benchmark. blogs.msdn.microsoft.com/stevengu/2019/…

"New State of the Art AI Optimizer: Rectified Adam (RAdam). Improve your AI accuracy instantly versus Adam, & why it works" It's been a long time since we've seen a new optimizer reliably beat the old favorites; this looks like a very encouraging approach! link.medium.com/iN3d4LMpbZ

Introducing unified Vision-Language Pre-training (VLP)! VLP is pre-trained on millions of image-text pairs and fine-tuned for captioning and VQA. We achieve SotA on COCO (C: 129), VQA 2.0 (Overall 71), all w/ a single model. lnkd.in/eARbUzU. Code: lnkd.in/eF9W3T5

Check out my latest article: UniLM (v1) is released! linkedin.com/pulse/unilm-v1… via LinkedIn

MT-DNN-SMART creates the new SOTA result on GLUE linkedin.com/pulse/mt-dnn-s… via LinkedIn

Check out my latest article: DIALOGPT : Large-Scale Generative Pre-training for Conversational Response Generation linkedin.com/pulse/dialogpt… via LinkedIn

AI has largely moved from symbol-based systems to artificial neural network–based models. TP-Transformer and TP-N2F show how a neurosymbolic approach that merges the two via neural symbols can enhance performance and interpretability: aka.ms/AA6s9xm #NeurIPS2019

Check out my latest article: MSR's new neurosymbolic models learn to encode and process neural symbols linkedin.com/pulse/msrs-new… via LinkedIn

Curious about how to advance & apply deep generative models in an era of large-scale pre-training? Learn from three new projects—Optimus for language modeling, FQ-GAN for image generation, and Prevalent for vision-language navigation—led by Chunyuan Li: aka.ms/AA84pcr

Microsoft’s DialoGPT helps build versatile, engaging and natural open-domain conversational agents. Explore the source code and trained model: aka.ms/AA88oin Decoding functionality from @HuggingFace: huggingface.co/microsoft/Dial…

In vision-and-language pretraining (VLP), objects can be used as anchor points to make aligning semantics between image-text pairs easier. Learn how Oscar, a novel VLP framework utilizing objects, sets new state of the art on six vision-and-language tasks: aka.ms/AA8frf5

The Microsoft AI model DeBERTa—being integrated into the next version of Microsoft Turing—set a new state of the art in natural language understanding. Learn how it uses 3 novel techniques to surpass human performance on and top the SuperGLUE leaderboard: aka.ms/AAarfkr

Microsoft researchers have created VinVL—a new object-attribute detection model that helps the Microsoft vision-language (VL) system top 3 VL leaderboards. Learn how the VinVL model enables much richer semantics and far greater image encoding: aka.ms/AAatqtv

Biomedicine is facing an information explosion. The Microsoft Biomedical Search prototype uses advances in natural language to enable scientists & clinicians search w/ natural language vs. keywords. More at aka.ms/AAbhb8c Microsoft Research National Library of Medicine #MicrosoftOCSO Computing Community Consortium (CCC)

We are releasing Focal Transformer! A new self-attention mechanism with fine-grain attention surrounding but gradually coarser-grain attention far away. By modeling the short and long-range visual dependencies in a focal manner, we are delivering new SoTA on COCO and ADE20K!

The Microsoft Turing model T-NLRv5 has reached new heights on SuperGLUE and GLUE leaderboards, reaffirming our commitment to bringing smarter, more responsible AI product experiences to our customers using state-of-the-art language models. msft.it/6004k0Ndr

After 1+ years of writing, the first draft of our book (w. Jianfeng Gao Paul Bennett Nick Craswell ) "Neural Approaches to Conversational Information Retrieval" is now up on ArXiv (arxiv.org/abs/2201.05176) for feedback. Let us know where you think we can improve the book!