10 Years of Neural Network Evolution in Yandex Search: From Image Search to Multimodal Neuro

In December 2014, we significantly upgraded Yandex's image search by adding neural networks. As a result, Yandex Search has become much better at finding semantically related images. To make it possible, we implemented convolutional neural networks, which had proven superior to other machine learning methods for image recognition just a few years prior.

Results powered by Yandex Search's early neural networks

This was not the first time Yandex used neural networks. In 2012, Yandex used a simple neural network to predict traffic conditions, and in 2013, neural networks were employed in speech recognition for our SpeechKit product.

But let's take a closer look at image search.

A game-changer came in 2015 when we introduced neural networks to improve our image search results for text queries dramatically. Before that, Yandex Search relied solely on the webpage text around the image to determine whether the image was relevant to a search. The new model mapped images and the query text into a common multimodal semantic space and compared them directly to determine if the image was relevant to the search.

In 2016, we introduced neural networks to our Palekh search algorithm to better understand user queries and improve website rankings. We developed a DSSM-like model to evaluate the semantic proximity of webpage titles and user queries. In 2017, with the Korolyov update, we extended the use of neural networks beyond titles to the full text of web pages. Yandex Search could now better understand long-tail unique queries and provide accurate results, even when there wasn't enough information to rank the responses.

In 2020, we pioneered the use of a large neural network called YATI (Yet Another Transformer with Improvements) to rank websites. While YATI was "just another transformer" in its basic architecture, it was fine-tuned and optimized to work in Yandex Search's runtime environment. This update delivered the most significant quality improvements to website ranking since the introduction of MatrixNet in 2009.

Yandex first used neural networks in machine translation in 2017. This allowed Yandex Translate to understand the context of phrases and sentences better, delivering more accurate translations. In 2021, Yandex Search and Yandex Browser gained the ability to translate videos and add voice-overs with multiple voices.

Yandex Search's mission is to help people accomplish more with ease. That's why Yandex Search doesn't just give you a list of websites; it can also provide a quick answer with links to reliable sources.

In 2024, we added a powerful new AI feature called Neuro. This feature enables our search to provide concise answers by summarizing the information it finds, allowing users to accomplish what they want much faster. The queries in Neuro are processed by a multimodal model, which aligns with today's standards. You can search with text, images, or a combination of both.

Contacts

Yandex Press Office
pr@yandex-team.com