Deepmind has developed a universal architecture for creating artificial intelligence
DeepMind laboratory specialists have created the universal architecture of PERCEIVER IO to process all types of input and output data.
To Tackle All the Challenges We Meet While Solving Intelligence, We NEED TOLS that AR AS ADAPTABLE as Possible. Announcing the Paper & Code for perceiver Io, An Architecture that Handles a Wide Range of Data and Tasks, All While Scaling GraceFully: Https: // T.CO/9uu9b6276q 1/4 pic.Twitter.COM/PO9JSHT7YL
– Deepmind (@deepmind) August 3, 2021
It is based on the original model of perceiver, presented in June 2021. She processes images, audio, video and their combinations, but is limited to tasks with simple weekends, such as classification.
To solve this problem, researchers have created a more common version of architecture – perceiver Io. It can issue a wide range of output data from the variety of incoming information, which makes it applicable to areas such as processing a natural language, computer vision and multimodal understanding.
Perceiver and perceiver IO are built on the architecture of transformers that work well for input data containing several thousand elements. However, according to researchers, images, audio and video may contain millions of such elements.
“With the help of the original perceiver, we solved the main problem of universal architecture: scaling transformers Emirates kicks for very large input data without the introduction of assumptions specific to the subject area,” the blog said.
Researchers also believe that perceiver Io can reach an unprecedented level of universality.
They published the source code of architecture on GitHub and hope that this will help researchers and practices develop applications without the need to spend resources on creating individual solutions using specialized systems.
Recall that at the end of July DeepMind introduced an extensive Xland game environment for teaching universal artificial intelligence agents.
In July, experts of the AI-laboratories gathered and published the most complete database of human protein structures created by Alphafold neural network.
Subscribe to FORKLOG news at Telegram: Forklog AI – all news from the world of AI!