Google Multitask Unified Model, MUM

[engine]

Google this week announced Multitask Unified Model (MUM), which is a new technology it's hoping will make complex queries. Mum is like BERT [webmasterworld.com], but much more powerful: Google says 1,000 time more powerful. Google calls it "MUM: A new AI milestone for understanding information"

MUM not only understands language, but also generates it. It’s trained across 75 different languages and many different tasks at once, allowing it to develop a more comprehensive understanding of information and world knowledge than previous models. And MUM is multimodal, so it understands information across text and images and, in the future, can expand to more modalities like video and audio.

MUM is multimodal, which means it can understand information from different formats like webpages, pictures and more, simultaneously. Eventually, you might be able to take a photo of your hiking boots and ask, “can I use these to hike Mt. Fuji?” MUM would understand the image and connect it with your question to let you know your boots would work just fine. It could then point you to a blog with a list of recommended gear.

It's not live yet, but i'm sure Google will be running tests over time, and it says over the coming months and years, so it must be confident this technology will work well.

[blog.google...]

Other stories that may be of interest Google : Rethinking Search: Making Experts out of Dilettantes [webmasterworld.com]

[JesterMagic]

Well Google's current AI is a joke. Just look at all the spam on the first page that a human reader can pick out in most cases very easy. They have a very long way to go.

Google's ultimate goal is to get every query down to that one answer (like when using home) and keep you on their own platform.

I much prefer a list of answers that I can peruse as a lot of times the searcher full intent can never be quantified by a short query sentence.

[iamlost]

Multitask Unified Model
MUM
aka the one ring (just kidding)

I’ve been interested in NLP for about a quarter century now. Beginning with GATE (General Architecture for Text Engineering) tool suite from the University of Sheffield and later the TensorFlo platform from Google.

A combination of mind stretching exercise and practical site enhancements. I’d got myself sort of comfortable with attention enhanced LSTM-RNN (Long short-term memory recurrent neural network) architecture with a side order of contextual bandits... and along came transformations. Oh boy. Oy vey. Oh what a pita.

Put simply transformations are matrix calculations while what came before were linear. And therein lies the basis of the ‘1000x more powerful’ claim. That each matrix position can have one or more weighted inputs acting upon it... ya. Vroom.

The Google examples used in the introduction of MUM were complex queries (almost certainly more likely to be voice queries than text, we’ve been too well ‘trained’ on existing text I/O SE limitations)

Given (1) my belief of the direction of Google’s search business model and (2) my severely restricted experience and understanding of transformations (I’m too old for this constant new math stuff) the following is my expectation of how Google will use the MU Model...

Currently the query result pulls a single ‘idea’ answer. One query one result. Granted, in some cases it may provide more than one because it’s confused, eg is it Washington the person, state, city.

With MUM the goal is to be able pull many results to a single query. The provided example, ‘what do I need to hike Mt Fuji in the fall that is different from hiking Mt Adams’ requires pulling differences data on each mountain, location seasonal weather, hiking gear/prep lists for each, etc.

I expect that when we see the first public iteration there will be a series of answers/snippets eg one for specific query point. Possibly in side to side comparison eg fall weather variation Adams and fall weather variation Fuji. Or it may default jump straight to providing only the weather variation difference.

I see the ultimate goal being able to keep all that initially front end series of results in the backend and be able to present a mashup single complex answer.

A single answer that while it was derived from a variety of sources (sites) is newly created content belonging solely to Google requiring no attribution and no link outs.

Google will have ingested the world’s information, organised the world’s information, and made it their own.

Google will be the source, the one source...

Note: matrices math is fun.
My brain hurts so good.
When is too early to add C2H6O to COFFEE?
Chemistry is yummy good...
Education.
STEM.
MuhaHaHA.