Single stream multi-level alignment for vision-language pretraining

Abstract

A method is provided for pretraining vision and language models that includes receiving image-text pairs, each including an image and a text describing the image. The method encodes an image into a set of feature vectors corresponding to input image patches and a CLS token which represents a global image feature. The method parses, by a text tokenizer, the text into a set of feature vectors as tokens for each word in the text. The method encodes the CLS token from the NN based visual encoder and the tokens from the text tokenizer into a set of features by a NN based text and multimodal encoder that shares weights for encoding both the CLS token and the tokens. The method accumulates the weights from multiple iterations as an exponential moving average of the weights during the pretraining until a predetermined error threshold is reduced to be under a threshold amount.