Timezone: »

VLMixer: Unpaired Vision-Language Pre-training via Cross-Modal CutMix
Teng Wang · Wenhao Jiang · Zhichao Lu · Feng Zheng · Ran Cheng · chengguo yin · Ping Luo

Tue Jul 19 11:35 AM -- 11:40 AM (PDT) @ Hall F

Recent vision-language pre-training (VLP) methods depend heavily on paired image-text datasets, which are annotated by enormous human labors, or crawled from the internet followed by elaborate data cleaning techniques. To reduce the dependency on well-aligned image-text pairs, it is promising to directly leverage the large-scale text-only and image-only corpora. In this paper, we propose a data augmentation method, namely cross-modal CutMix (CMC), for cross-modal alignment learning in unpaired VLP. Specifically, CMC transforms natural sentences in the textual view into a multi-modal view, where visually-grounded words in the sentence are randomly replaced by diverse image patches with similar semantics. The CMC has several appealing proprieties. First, it increases the data diversity while keeping semantic meaning nearly unchanged, which could tackle the problem where the aligned data are scarce; Second, by attaching cross-modal noise on uni-modal data, it guides the model to learn token-level interactions across modalities for better denoising. Furthermore, we present a new unpaired VLP method (dubbed as VLMixer) that novelly integrates CMC with contrastive learning to pull together the uni-modal view and the multi-modal view, for better instance-level alignments between different modalities. Extensive experiments on five downstream tasks verify the effectiveness of CMC and show that VLMixer could surpass previous state-of-the-art unpaired VLP methods.

Author Information

Teng Wang (Southern University of Science and Technology)
Wenhao Jiang (Tencent)
Zhichao Lu (Southern University of Science and Technology)
Feng Zheng (SUSTech)
Ran Cheng (Southern University of Science and Technology)
Ran Cheng

Dr. Ran Cheng, the founder of the Evolving Machine Intelligence (EMI) Group, is currently a tenured Associate Professor with the Southern University of Science and Technology (SUSTech), China. He received the PhD degree in computer science from the University of Surrey, UK, in 2016. His research interests mainly fall into the interdisciplinary fields across evolutionary computation and other major AI branches such as statistical learning and deep learning, to provide end-to-end solutions to optimization & modeling in scientific research and engineering related applications. He is the Founding Chair of IEEE Computational Intelligence Society (CIS) Shenzhen Chapter and IEEE Symposium on Model Based Evolutionary Algorithms (IEEE MBEA). He serves as an Associated Editor/Editorial Board Member for serveral jounrlas, including: IEEE Transactions on Evolutionary Computation, IEEE Transactions on Cognitive and Developmental Systems, IEEE Transactions on Artificial Intelligence, etc. He is the recipient of the IEEE Transactions on Evolutionary Computation Outstanding Paper Awards (2018, 2021), the IEEE CIS Outstanding PhD Dissertation Award (2019), the IEEE Computational Intelligence Magazine Outstanding Paper Award (2020). He is a Senior Member of IEEE.

chengguo yin (tencent)
Ping Luo (The University of Hong Kong)

Related Events (a corresponding poster, oral, or spotlight)

More from the Same Authors