Microsoft announces launch of Turing Bletchley v3 visual-language model for Bing Image Search

2023-08-31

Microsoft has officially announced the third version of its Turing Bletchley multilingual vision language base model. It is now being applied to multiple Microsoft products, including the Bing search engine, to improve image search. Microsoft launched the first version of the Turing Bletchley model in November 2021. According to Microsoft's official blog on Bing, the third version of the model is being tested in the fall of 2022 and will be added to Bing and other products. The model utilizes text and image inputs to find the content users are searching for on the Microsoft Bing search engine. The goal is to make the model as close as possible to, for example, matching the description "dog eating ice cream" with images of a dog eating ice cream in the search results. Part of establishing these connections in Turing Bletchley v3 involves comprehensive adjustments to the model. Microsoft explains: "Given an image and a caption describing the image, certain words in the caption are masked. Then a neural network is trained to predict the hidden words based on the image and text. This task can also be reversed, replacing pixels with words. This masked training, combined with powerful pre-training models based on large-scale Transformer models, allows for fine-tuning on diverse training sets for various downstream tasks." In addition to image search for Bing, the new Turing Bletchley v3 model is also used for content moderation in Xbox game services. It helps the team identify images and videos uploaded to Xbox platform user profiles, determine if they violate the company's community guidelines on the Xbox platform, and if they contain inappropriate content.

RECENT AI NEWS

RECENT AI TOOLS