Advances in artificial intelligence (AI) are not stopping. Today, algorithms are behind many everyday tasks. From taking a good photo with your mobile to asking a virtual assistant for something. Google has been using it for 20 years and is now ready to take advantage of all that experience in a new generation architecture called Pathways .
Google began its journey with artificial intelligence in 2001. At that time, a machine learning technique made it possible to provide accurate results, even if users misspelled the search term. That was the starting point to roll out to almost all of its services and applications. But during all this time the same training method was used.
In the face of every problem, for example spam, a new model of artificial intelligence is trained from scratch. That is, instead of extending what has been learned in previous solutions, new very specific solutions are created , which require resources and time. While this has been tremendously useful so far, it is not entirely efficient and requires evolution.
Pathways, ready for today's challenges
Imagine if, every time you learned a new skill (jumping rope, for example), you forgot everything you had learned (how to balance, how to jump, how to coordinate the movement of your hands) and began to learn each new skill of the rope. nothing “, exemplifies Google in reference to current methods to train artificial intelligence.
The result of this? Thousands of wildly specific models for solving problems. But this could be coming to an end with Pathways architecture. This is capable of training “ a single model to do thousands or millions of things ” compared to the current highly individualized approach. And since you don't have to start from scratch every time, it would be much more efficient than the old methods.
Google notes that Pathways can “leverage and combine your existing skills to learn new tasks more quickly and effectively.” The company thinks of this artificial intelligence in architecture as the brain of mammals. That is, it can handle many different tasks. From the simplest to the most complex.
Google and an artificial intelligence with multiple 'senses'
Credit: Google Current artificial intelligence models, in addition to being tremendously specific, focus on only one “sense.” That is, they process a single type of information at a time. They can take text, images or voice, but individually, something that limits their ability to face certain tasks and presents biases .
However, Pathways also seeks to move beyond the limits of current models. To do this, we are working on a multimodal model that could allow simultaneous visual and auditory processing . So, Google's system could identify a leopard if you see a video of the animal running, listen to it, read the word leopard, or someone says “leopard”.
With advances like the one presented today, artificial intelligence could help tackle challenges like disease, inequality and climate change, according to Google. The truth is that the company is also working hard on its Unified Multitasking Model (MUM), which has already shown that it is capable of being very useful to improve search results.
