Google Pathways is a new, next-generation AI technology that allows neural networks to perform up to millions of tasks.
Current AI models are often trained to perform a single task. With the development of Pathways, the tech giant aims to take a step towards hyperintelligent AI models that can perform multiple tasks.
Neural networks must be given multiple ‘skills’ for this purpose, Google indicates. By combining different skills, multiple tasks can be performed. Additionally, Pathways promises to enable AI models to use different types of data. Neural networks should soon be able to extract data from other sources than text, audio and video, enabling neural networks to make better and more accurate decisions. Moreover, present neural network errors and prejudices are likely to be more effectively prevented.
Pathways also aims to enable neural networks to work more efficiently. As it stands, AI models activate all of their artificial neurons when they make calculations — even the neurons they are not using. The latter has major implications for the required underlying infrastructure. With Pathways, Google wants the models to activate only those parts of the neural network that are needed to perform a specific task. Doing so should make more (infrastructure) capacity available for learning multiple tasks, allow for faster processing, and lighten energy consumption.
Additionally, Google has been considering what the new AI technology and the performance of million tasks by AI models can be used for. Among other things, the company notes autonomous vehicles. These vehicles must perform multiple computational tasks for turning data from sensors into driving decisions. Usually, each computational task is performed through a separate neural network. By replacing these different AI models with a single model that can handle all tasks, efficiency and performance improve.
The tech giant’s public cloud is another use case. Google states the potential for strengthening the current cloud-based managed AI services portfolio.
Google is currently applying Pathways in several projects. Google Switch Transformer, a natural language processing (NLP) model based on Pathways, functions on a relatively low number of artificial neurons for calculations.