Microsoft Bing introduces ‘Make Every Feature Binary’

Microsoft revealed their latest AI model, a large-scale sparse model that supports our production Transformer models, which they are calling MEB (Make Every Feature Binary). Microsoft claims that this makes Bing search results more relevant.

In fact, MEB is currently in production for 100% of Bing searches across all regions and languages. However, Microsoft stated that “to make search more accurate and dynamic, MEB better harnesses the power of large data and allows for an input feature space with over 200 billion binary features that reflect the subtle relationships between search queries and documents.”

MEB correlates each fact to a feature, according to Microsoft, “so it can give weights that distinguish between the capacity to fly in, say, a penguin and a puffin, and it can do this for each of the qualities that make a bird—or any entity or item for that matter—singular.”

MEB, how clever are you? It did, however, go through extensive training. MEB has been trained with over 500 billion query/document pairings gleaned from three years of Bing searches. According to Microsoft, the input feature space has more than 200 billion binary features.

According to Microsoft, the MEB model also displays intriguing possibilities for learning beyond semantic connections. When investigating the top MEB features, Microsoft discovered that it can understand hidden intentions between query and document.

According to Microsoft, which released the following data points, a nearly 2% boost in click-through rate (CTR) on top search results. Those results that are found “above the fold” without having to scroll down.