Microsoft Announces The Release Of Their Two Billion Parameter Latest Vision-Language AI Model Called BEiT-3

Microsoft’s Natural Language Computing (NLC) team recently introduced their latest vision-language AI model, BEiT-3, a Bidirectional Encoder representation from Image Transformers with 1.9 billion parameters. 

BEiT-3’s central idea is to treat images as if they were written in a different language (which the authors refer to as “Imglish”), allowing the model to be pretrained with simply the masked language modeling (MLM) objective. Because of its unified architecture, BEiT-3 can support a wide variety of downstream tasks. In evaluation experiments, the …

ai model computervision language language ai microsoft release vision