Meta-Transformer: A Unified Framework for Multimodal Learning — Unified Multimodal Learning. Meta-Transformer utilizes the same backbone to encode natural languages, images, point clouds, audio, video, infrared, hyperspectral, X-ray, time-series, tabular, and graph data. It reveals the potential of transformer architectures for universal perception.
Multimodal learning involves utilizing data from various modalities to improve model capacity. Despite the years of development in this field, it remains challenging to devise a unified framework for processing natural language, 2D images, 3D point clouds, and audio spectrograms due to crucial gaps among these different modalities. This study proposes a novel approach that demonstrates a network with frozen parameters can encode the data from the aforementioned four modalities and achieve favorable performance, resulting in a unified framework called Meta Transformer. Using this framework, the raw input data from various modalities are converted to a shared token space, allowing a subsequent encoder with frozen parameters to extract high-level semantic features of the input data. Composed of three main components: a unified data tokenizer, a modality-shared encoder, and task-specific heads for downstream tasks …