DeepMind Releases Perceiver IO as Open-Source Alternative to Transformers
DeepMind open-sources Perceiver IO, a versatile deep learning model that handles text, images, and sound without the heavy computational costs of traditional transformers.
DeepMind releases Perceiver IO as an open-source deep learning model designed to handle various types of data inputs and outputs. This general-purpose architecture serves as a viable replacement for traditional transformers by using attention mechanisms to map inputs into a latent representation space.
Unlike standard transformers, Perceiver IO processes longer input sequences without suffering from quadratic compute and memory losses. The model demonstrates remarkable versatility by seamlessly working with different data types, including text, images, and sound, and decoding the output into any required format.
DeepMind researchers test Perceiver IO across multiple tasks and find that it performs comparably to existing models in language understanding, image classification, and video game playing. The model also achieves high results on the Sintel optical flow benchmark, proving its potential to unify multi-modal data processing under a single, efficient architecture.