This 2021 edition of Google I/O, which takes place in streaming from May 18 to 20, is, as we could expect, the occasion of a multitude of interesting announcements in terms of artificial intelligence. On the hardware front, we previously mentioned a new version of the TPU . A second announcement could have important repercussions: that of Multitask Unified Model(MUM). MUM is an automated language processing model that pushes the analysis even further than BERT.

Artificial intelligence technology to process certain queries faster

MUM aims to help the user to perform complex tasks. Fourth of its kind developed by Google after Hummingbird, Rankbrain and BERT, the model is based, like BERT(Bidirectional Encoder Representations from Transformers), on a Transformer architecture (an automated language processing technology to be followed closely, presented by the ActuIA scientific committee in the ActuIA magazine N°3).

MUM would be 1000 times more powerful than BERT.

Trained in 75 languages and to perform many tasks at once, it has been developed to understand much more information than previous models of its kind. Google hopes that with this new model, fewer searches will be needed in the future to solve complex problems.

A multimodal technology

MUM is multimodal, which means that it can simultaneously understand information from different formats: web pages, text, images, photos, videos, etc. Below is an example of how MUM is used. A question is asked by the user: “Can I use these shoes to hike on Mount Fuji”. The person also adds a picture of the shoes in question. MUM analyzes the picture and relates it to the question to answer it and redirect you to a suitable site if the boots are not suitable or tell you that your boots are perfectly suitable by providing you with an article that explains why.

Moreover, let’s imagine that the user is not Japanese, but French. If the person had done a search with the Google search engine, they would probably never have come across Japanese content that might be interesting. MUM will have the ability to search for information in all languages and provide the user with the best results that can be found for a search in Japanese, for example.

Internal pilot projects have been launched with MUM and the company has been pleased with the results and the potential benefits that the tool brings to its entire product line. Google has applied all of its latest research to develop MUM while keeping in mind to reduce its carbon footprint. The brand states that:

“MUM is an important step towards a future where Google can understand the full range of ways people use to communicate and interpret information naturally.”

The nature of Google’s search engine continues to evolve

Google was historically a simple search engine, consisting of a web form and a database of previously indexed web pages. Its role was to suggest that we visit the most relevant sites.

The improvement of its understanding of language has allowed it to better target relevant content by interpreting the user’s intention, but above all to extract the content of web pages to display it in the results page, thus gradually transforming the search engine into an assistant. This evolution is not without questioning more and more website editors, some of whom feel they have been reduced to a source of knowledge for the engine. These publishers could of course refuse to have their content indexed on Google, but they know they are dependent on the engine. It will be very interesting to observe the evolution of the situation in the medium and long term. It’s a safe bet that the next obstacle to this trend will not be technical.

Translated from Google I/O : annonce de la technologie d’intelligence artificielle MUM, plus puissante que BERT