DeepMind Releases Perceiver IO, a General-Purpose Deep Learning Model
DeepMind open-sources Perceiver IO, a versatile architecture that processes multiple data types without quadratic computational costs. The model matches or exceeds specialized baselines across various AI benchmarks.
DeepMind releases Perceiver IO, a general-purpose deep-learning model architecture that processes many different types of inputs and outputs. This open-source model serves as a "drop-in" replacement for Transformers, performing as well or better than baseline models without relying on domain-specific assumptions.
The architecture builds on the Transformer model by using attention to map inputs into a latent representation space. Unlike standard Transformers, Perceiver IO handles longer input sequences without incurring quadratic compute and memory costs, and it accepts simultaneous inputs of multiple data types including text, images, and sound.
In extensive experiments, the DeepMind team evaluates Perceiver IO on language understanding, image classification, optical flow, and video-game playing. The model performs comparably to state-of-the-art models on most tasks and achieves new state-of-the-art results on the Sintel optical flow benchmark, offering researchers a versatile tool to avoid building custom specialized solutions.