Pathways, Google's next-generation multitasking AI
On October 28, Google introduced Pathways, a new AI architecture that can perform multiple tasks at once.
Google unveiled a new model of artificial intelligence capable of performing a much wider range of tasks than the specialized models currently in use. The Pathways AI model involves combining the specialized systems currently in use into a multi-modal, universal system. This flexibility could allow AI to function like a human brain, with the advantages and disadvantages of a neural network.
The standard AI model is trained in one way to perform a single task. Mix enough of these algorithms together and you get AI engines for voice assistants and other software that seem monolithic to users who can't see the complex fabric within them. Google claims that Pathways can combine all of these individual algorithms into a universal neural network capable of performing various tasks and learning. This is in stark contrast to starting from scratch with each new function that AI trains, as if it has forgotten everything, and requires a lot of extra time and data from engineers.
"This is how or about how we train most machine learning models today. Instead of extending existing models to train new tasks, we train each new model from scratch to do one thing and one thing only (or sometimes we specialize a generic model for a specific task). We end up developing thousands of templates for thousands of individual tasks", explains Jeff Dean, senior vice president of Google Research, on the Pathways blog. "Instead, we'd like to train a single model that can not only handle many individual tasks, but also use and combine existing skills to learn new tasks faster and more efficiently."
The AI, based on the Pathways model, will retain previous learning in its neural network and use it in the future. The AI can apply its knowledge from existing skills to learn new skills by weaving together points where disparate concepts have similarities. Dean compared contextual memory to how the mammalian brain works. He suggested that an AI trained to use aerial images to estimate elevation could extend that knowledge to predict how a flood might cross a mountain valley, without having to specifically train a new algorithm as is done today.
To develop flexible responses to models using Pathway AIreport, Google is applying a similar principle to make Pathways multimodal in real-world data collection. Today, a typical neural network can process text, audio or video, but not all three. Google believes that Pathways is advanced enough to collect all three types of data and understand how they interact. These three types of data enable decisions to be made in the Pathways neural network. Data collection can be translated from one format to another, so Pathways can be used alone or as a complement to existing systems that struggle to collect enough data without opening up to new modes of communication.
"Humans rely on multiple senses to perceive the world. This is very different from the way modern artificial intelligence systems digest information. Most modern models only process one type of information at a time. They can perceive text, images or speech, but usually not all three at once", explains Mr. Dean. "That's why we created Pathways", said Dean. "Pathways will enable a single AI system to summarize thousands or millions of tasks, understand different types of data, and do so with astonishing efficiency, allowing us to move from an era of single-purpose models that simply recognize patterns, to an era in which more versatile intelligent systems reflect a deeper understanding of our world and can adapt to new needs."