Google’s New MUM (Multitask Unified Model) System
Google's MUM (Multitask Unified Model) system is the new way the search engine understands what people are searching for. Google has been gradually moving away from traditional keyword research to focus on the bigger picture of what searchers are asking, their situation, and what they need to do next. This new algorithm uses machine learning to understand all of this.
When we talk to each other, we convey information using multiple modes – written, spoken, and sometimes even different visual representations. That's why MUM is so powerful – it understands information across texts and images and is the product of over two years of research and engineering from the Google Brain team. In this post, we'll take a deeper look at what MUM is and how it works.
What Exactly is Google's MUM, and What Does it Do?
Recently, Google revealed its "Multitask Unified Model," a new AI system that would boost its capabilities in areas like image recognition and language processing. The search giant published an article on the technology alongside a video that shows how MUM works.
The MUM system is a significant advance for Google, which has been working in recent years to build its own AI systems in-house, and it stands to improve their own AI offerings significantly.
As Google describes it, the MUM system is "a machine learning architecture for end-to-end learning of complex multimodal tasks" — in other words, AI that can process different inputs, such as images, sounds, text and speech, simultaneously.
MUM builds on that technology and adds the ability to process multiple inputs at the same time. As Google puts it, MUM "employs a set of tightly coupled modules that learn concurrently, sharing weights and activations across their outputs."
How Does MUM Relate to BERT and How Powerful Is It?
The changes coming from Google's new MUM are significant. They are significant because, for the first time, Google has created a single neural network capable of both natural language processing (NLP) and machine translation (MT).
Its predecessor, BERT, which stands for Bidirectional Encoder Representations from Transformers, is a neural network designed to perform NLP, and it's already one of the company's most powerful NLP tools. Still, BERT isn't as great for MT.
On the other hand, MUM is explicitly designed for MT, and it's capable of performing translation in both directions. The potential of this is enormous, but one big question remains: How does MUM relate to BERT?
BERT's purpose is similar to that of MUM, but MUM is much more powerful. The BERT project was initially developed out of Google's work on machine translation, but researchers for NLP later repurposed it.
MUM was developed independently, and its specific purpose goes beyond pure MT. For example, it could be used to transcribe video, which means it could be extremely valuable for Google's new Assistant and Google Search.
With MUM, Google Can Understand Context Across Different Mediums
MUM is a set of AI models built on top of the Google Assistant framework that understands context across different mediums. This lets Google's AI better understand the context of what it sees, understands, hears, and interacts with.
For example, MUM can understand what you're trying to communicate and help you do that, whether through Google Now, the Google Assistant, or smart speakers like the Google Home. Also, it will eventually be included in all Google searches.
At a high level, MUM is about understanding the user. Google wants to understand the user's intent, imagination, and context.
How MUM Will Affect the Industry and SEO in the Future?
While Google hasn't released all of the details of MUM, we know that it focuses on three major factors:
- Content Quality - Google's algorithm has traditionally tried to rank sites based on a combination of factors, like how many backlinks the site has and how well it's written. But MUM focuses heavily on content quality, especially as it applies to the searchers' intent. So if you're creating content, focus on quality and don't try to game the system with junky or rushed work.
- A Search User Experience - Google's algorithms have always focused on whether or not a site provides a quality user experience. MUM takes this further by looking at your site's overall layout, functionality, and navigational structure. If Google thinks your site is challenging to use for its searchers, it will likely rank you lower.
- Local Intent - Google's algorithms have always tried to rank sites based on where they're being generated. So, for example, if you're generating content for local traffic, your site will be ranked higher. Currently, local, in this case, does not mean at the city or street level but at the country level.
MUM will likely continue this trend but takes it a bit further. Google will likely rank sites based on where they are being generated or where their servers live. MUM suggests that if your site is hosted in China, it will assume it's more likely to be trustworthy for Chinese users than a site that is hosted in the U.S.
This is a shift that will likely affect many SEOs because Google's algorithms are usually very U.S.-centric. But MUM is likely to be a game-changer even for those who aren't trying to generate local country traffic because Google's algorithms traditionally focus on on-site quality, not geography.
Key Google MUM Takeaways
The MUM system, in theory, will dramatically improve the speed at which search results are returned. This is a big deal, as Google constantly wants to deliver the fastest possible results to its users.
The concepts behind MUM are nothing new, but Google's implementation is. The company's engineers have integrated machine learning directly into the search engine, which means it's no longer a passive system with predictive capabilities. Essentially, the search engine will be able to anticipate what a user will be looking for and proactively return relevant results.
This is a big deal, but it's not without its challenges. A secondary benefit of the MUM system is that search results will be more accurate, leading to less irrelevant and (potentially) misleading results. However, that comes with a cost for Google, as the search engine has to consider every future query, every search and return the best possible results. It's a massive undertaking, but if anyone has the resources to make this happen, Google does.
It is expected that the MUM system should be rolled out to all users by the end of the year, and it's already helping to improve search performance.