Existing Vision-Language Pretraining (VLP) methods have achieved remarkable improvements across a variety of vision-language tasks, confirming their effectiveness in capturing coarse-grained semantic correlations. However, their capability for fine-grained understanding, which is critical for many nuanced vision-language applications, remains limited. Prevailing VLP models often overlook the intricate distinctions in expressing different modal features and typically depend on the similarity of holistic features for cross-modal interactions. Moreover, these models directly align and integrate features from different modalities, focusing more on coarse-grained general representations, thus failing to capture the nuanced differences necessary for tasks demanding a more detailed perception. In response to these limitations, we introduce Negative Augmented Samples(NAS), a refined vision-language pretraining model that innovatively incorporates NAS to specifically address the challenge of fine-grained understanding. NAS utilizes a Visual Dictionary(VD) as a semantic bridge between visual and linguistic domains. Additionally, it employs a Negative Visual Augmentation(NVA) method based on the VD to generate challenging negative image samples. These samples deviate from positive samples exclusively at the token level, thereby necessitating that the model discerns the subtle disparities between positive and negative samples with greater precision. Comprehensive experiments validate the efficacy of NAS components and underscore its potential to enhance fine-grained vision-language comprehension.

Despite the significant success of Large Vision-Language models(LVLMs), these models still suffer hallucinations when describing images, generating answers that include non-factual objects. It is reported that these models tend to overfocus on certain irrelevant image tokens that do not contain critical information for answering the question and distort the output. To address this, we propose an Instruction-Aligned Visual Attention(IAVA) approach, which identifies irrelevant tokens by comparing changes in attention weights under two different instructions. By applying contrastive decoding, we dynamically adjust the logits generated from original image tokens and irrelevant image tokens, reducing the model’s over-attention to irrelevant information. The experimental results demonstrate that IAVA consistently outperforms existing decoding techniques on benchmarks such as MME, POPE, and TextVQA in mitigating object hallucinations. Our IAVA approach is available online at https://github.com/Lee-lab558/IAVA.

Yufei Ma, Zihan Liang, Huangyu Dai, Ben Chen, Dehong Gao, Zhuoran Ran, Wang Zihan, Linbo Jin, Wen Jiang, Guannan Zhang, Xiaoyan Cai, Libin Yang. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing. 2024.

We present a new vision-language (VL) pre-training model dubbed Kaleido-BERT , which introduces a novel kaleido strategy for fashion cross-modality representations from transformers. In contrast to random masking strategy of recent VL models, we design alignment guided masking to jointly focus more on image-text semantic relations. To this end, we carry out five novel tasks, i.e., rotation, jigsaw, camouflage, grey-to-color, and blank-to-color for self-supervised VL pre-training at patches of different scale. Kaleido-BERT is conceptually simple and easy to extend to the existing BERT framework, it attains state-of-the-art results by large margins on four downstream tasks, including text retrieval (R@1: 4.03% absolute improvement), image retrieval (R@1: 7.13% abs imv.), category recognition (ACC: 3.28% abs imv.), and fashion captioning (Bleu4: 1.2 abs imv.). We validate the efficiency of Kaleido-BERT on a wide range of e-commerical websites, demonstrating its broader potential in real-world applications.

Question Answering (QA) systems have been extensively studied in both academia and the research community due to their wide real-world applications. When building such industrial-scale QA applications, we are facing two prominent challenges, i.e., i) lacking a sufficient amount of training data to learn an accurate model and ii) requiring high inference speed for online model serving. There are generally two ways to mitigate the above-mentioned problems. One is to adopt transfer learning to leverage information from other domains; the other is to distill the “dark knowledge” from a large teacher model to small student models. The former usually employs parameter sharing mechanisms for knowledge transfer, but does not utilize the “dark knowledge” of pre-trained large models. The latter usually does not consider the cross-domain information from other domains. We argue that these two types of methods can be complementary to each other. Hence in this work, we provide a new perspective on the potential of the teacher-student paradigm facilitating cross-domain transfer learning, where the teacher and student tasks belong to heterogeneous domains, with the goal to improve the student model’s performance in the target domain. Our framework considers the “dark knowledge” learned from large teacher models and also leverages the adaptive hints to alleviate the domain differences between teacher and student models. Extensive experiments have been conducted on two text matching tasks for retrieval-based QA systems. Results show the proposed method has better performance than the competing methods including the existing state-of-the-art transfer learning methods. We have also deployed our method in an online production system and observed significant improvements compared to the existing approaches in terms of both accuracy and cross-domain robustness.

In this paper, we address the text and image matching in cross-modal retrieval of the fashion industry. Different from the matching in the general domain, the fashion matching is required to pay much more attention to the fine-grained information in the fashion images and texts. Pioneer approaches detect the region of interests (i.e., RoIs) from images and use the RoI embeddings as image representations. In general, RoIs tend to represent the "object-level" information in the fashion images, while fashion texts are prone to describe more detailed information, e.g. styles, attributes. RoIs are thus not fine-grained enough for fashion text and image matching. To this end, we propose FashionBERT, which leverages patches as image features. With the pre-trained BERT model as the backbone network, FashionBERT learns high level representations of texts and images. Meanwhile, we propose an adaptive loss to trade off multitask learning in the FashionBERT modeling. Two tasks (i.e., text and image matching and cross-modal retrieval) are incorporated to evaluate FashionBERT. On the public dataset, experiments demonstrate FashionBERT achieves significant improvements in performances than the baseline and state-of-the-art approaches. In practice, FashionBERT is applied in a concrete cross-modal retrieval application. We provide the detailed matching performance and inference efficiency analysis.