The current status and prospects of transformer in multimodality

Abstract

At present, the attention mechanism represented by transformer has greatly promoted the development of natural language processing (NLP) and image processing (CV). However, in the multimodal field, the application of attention mechanism still mainly focuses on extracting the features of different types of data, and then fusing these features (such as text and image). With the increasing scale of the model and the instability of the Internet data, feature fusion has been difficult to solve the growing variety of multimodal problems for us, and the multimodal field has always lacked a model that can uniformly handle all types of data. In this paper, we first take the CV and NLP fields as examples to review various derived models of transformer. Then, based on the mechanism of word embedding and image embedding, we discuss how embedding with different granularity is handled uniformly under the attention mechanism in multimodal scenes. Further, we reveal that this mechanism will not only be limited to CV and NLP, but the real unified model will be able to handle tasks across data types through pre-training and fine tuning. Finally, on the specific implementation of the unified model, this paper lists several cases, and analyzes the valuable research directions in related fields.