Pathways: A next generation AI architecture that Google is working on

In the past two decades, innovation in the field of computer science has reached new heights, especially in the sphere of Artificial Intelligence (AI). One such innovation is the creation of pathways, a new-age AI architecture.  

What is Pathways? 
Pathways is a next generation AI architecture that are currently being developed by a team of Google researchers. This new technology utilises AI to address discrepancies of the existing systems and converts them into visible and upgraded changes.  

What challenges can Pathways address? 
In order to understand the process in details, let us discuss some of the current challenges and shortcomings of AI and how Pathways can bring about a significant improvement in those areas.  

Pathways will enable a single AI model to perform multiple tasks. The traditional AI models are built in such a way that they can accomplish only one objective, i.e., they are custom-made to perform a single task. Pathways will totally change this approach. It will enable experts to train a single AI model to accomplish more than one to unlimited goals. Not only will this speed up the process of completing simultaneous tasks without any pause, it will also reduce the amount of data that needs to be stored for each different model. Google team is working in such a way that this new model will help in carrying out separate and complex tasks and also learn the existing tasks and continue with them, in a quicker and more effective manner. 

Pathways will focus on multiple senses rather than the traditional single sense. The traditional AI model is trained to utilise a single sense at once, i.e., text or images or speech, but not three of them at once. Pathways will upgrade this feature. Just like humans use multiple senses together to perceive the world, Pathways will follow a similar pattern and retain all the information at once and use the senses incorporated in it simultaneously, making multi-tasking spontaneous and possible. Pathways will create, what is called multimodal AI model comprising of vision, auditory and language understanding all at once. For example, it will be able to detect the word “tiger,” will also respond to the sound of a “tiger” and capture the image or video of a “tiger” and analyse all of it as a single concept “tiger.” This is not only more insightful but will also help avoid mistakes.  

Pathways will help AI models complete one task using only a part of their neural network. Most traditional AI models are dense, i.e., their entire neural network is trained to accomplish one task, regardless of its difficulty level. This is almost like humans using their whole brain to solve a single problem. The only thing is, we hardly ever do that. Instead, we use certain specific neurons to perform a certain task, while the rest of the brain is at rest. Pathways will exactly follow this system. This new AI architecture will train the model in such a way, that it will be sparse active, i.e., only a certain part of the model will be called during a specific action, and not the whole of it. This will enhance efficiency, accuracy and speed, save time and energy and make the process dynamic. In fact, Pathways is being made in such a way that it will direct individual networks of the models to learn distinct functions.