Vision-Language Pre-Training with Triple Contrastive Learning

Multimodal representation learning has shown promising improvements on various vision-language tasks (e.g., image-text retrieval, visual question answering, etc) and has significantly advanced the development of multimedia information systems. Most existing methods excel at building global-level alignment between vision and language while lacking effective fine-grained image-text interaction. In this paper, we propose a jointly masked multimodal modeling method to learn fine-grained multimodal representations. Our method performs joint masking on image-text input and integrates both implicit and explicit targets for the masked signals to recover. The implicit target provides a unified and debiased objective for vision and language, where the model predicts latent multimodal representations of the unmasked input. The explicit target further enriches the multimodal representations by recovering high-level and semantically meaningful information: momentum visual features of image patches and concepts of word tokens. Through such a masked modeling process, our model not only learns fine-grained multimodal interaction, but also avoids the semantic gap between high-level representations and low-or mid-level prediction targets (e.g., image pixels, discrete vision tokens), thus producing semantically rich multimodal representations that perform well on both zero-shot and fine-tuned settings. Our pre-trained model (named MAMO) achieves state-of-the-art performance on various downstream vision-language tasks, including image-text retrieval, visual question answering, visual reasoning, and weakly-supervised visual grounding.

Data augmentation is a necessity to enhance data efficiency in deep learning. For vision-language pre-training, data is only augmented either for images or for text in previous works. In this paper, we present MixGen: a joint data augmentation for vision-language representation learning to further improve data efficiency. It generates new image-text pairs with semantic relationships preserved by interpolating images and concatenating text. It's simple, and can be plug-and-played into existing pipelines. We evaluate MixGen on four architectures, including CLIP, ViLT, ALBEF and TCL, across five downstream vision-language tasks to show its versatility and effectiveness. For example, adding MixGen in ALBEF pre-training leads to absolute performance improvements on downstream tasks: image-text retrieval (+6.2% on COCO fine-tuned and +5.3% on Flicker30K zero-shot), visual grounding (+0.9% on Re-fCOCO+), visual reasoning (+0.9% on NLVR <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> ), visual question answering (+0.3% on VQA2.0), and visual entail-ment (+0.4% on SNLI-VE).

The ability to model intra-modal and inter-modal interactions is fundamental in multimodal machine learning. The current state-of-the-art models usually adopt deep learning models with fixed structures. They can achieve exceptional performances on specific tasks, but face a particularly challenging problem of modality mismatch because of diversity of input modalities and their fixed structures. In this paper, we present Switch-BERT for joint vision and language representation learning to address this problem. Switch-BERT extends BERT architecture by introducing learnable layer-wise and cross-layer interactions. It learns to optimize attention from a set of attention modes representing these interactions. One specific property of the model is that it learns to attend outputs from various depths, therefore mitigates the modality mismatch problem. We present extensive experiments on visual question answering, image-text retrieval and referring expression comprehension experiments. Results confirm that, whereas alternative architectures including ViLBERT and UNITER may excel in particular tasks, Switch-BERT can consistently achieve better or comparable performances than the current state-of-the-art models in these tasks. Ablation studies indicate that the proposed model achieves superior performances due to its ability in learning task-specific multimodal interactions.KeywordsMultimodal interactionsCross-layer interactionSwitch attention

Deep learning methods have revolutionized speech recognition, image recognition, and natural language processing since 2010. Each of these tasks involves a single modality in their input signals. However, many applications in the artificial intelligence field involve multiple modalities. Therefore, it is of broad interest to study the more difficult and complex problem of modeling and learning across multiple modalities. In this paper, we provide a technical review of available models and learning methods for multimodal intelligence. The main focus of this review is the combination of vision and natural language modalities, which has become an important topic in both the computer vision and natural language processing research communities. This review provides a comprehensive analysis of recent works on multimodal deep learning from three perspectives: learning multimodal representations, fusing multimodal signals at various levels, and multimodal applications. Regarding multimodal representation learning, we review the key concepts of embedding, which unify multimodal signals into a single vector space and thereby enable cross-modality signal processing. We also review the properties of many types of embeddings that are constructed and learned for general downstream tasks. Regarding multimodal fusion, this review focuses on special architectures for the integration of representations of unimodal signals for a particular task. Regarding applications, selected areas of a broad interest in the current literature are covered, including image-to-text caption generation, text-to-image generation, and visual question answering. We believe that this review will facilitate future studies in the emerging field of multimodal intelligence for related communities.

Visual representation learning plays a key role in solving medical computer vision tasks. Recent advances in the literature often rely on vision-language models aiming to learn the representation of medical images from the supervision of paired captions in a label-free manner. The training of such models is however very data/time intensive and the alignment strategies involved in the contrastive loss functions may not capture the full richness of information carried by inter-data relationships. We assume here that considering expert knowledge from the medical domain can provide solutions to these problems during model optimization. To this end, we propose a novel knowledge-augmented vision-language contrastive representation learning framework consisting of the following steps: (1) Modeling the hierarchical relationships between various medical concepts using expert knowledge and medical images in a dataset through a knowledge graph, followed by translating each node into a knowledge embedding; And (2) integrating knowledge embeddings into a vision-language contrastive learning framework, either by introducing an additional alignment loss between visual and knowledge embeddings or by relaxing binary constraints of vision-language alignment using knowledge embeddings. Our results demonstrate that the proposed solution achieves competitive performances against state-of-the-art approaches for downstream tasks while requiring significantly less training data. Our code is available at https://github.com/Wxy-24/KL-CVR . • Rich semantics can be developed in representation learning by leveraging. • The inter-data relationships can be modelled using external knowledge. • Additional cross-modal alignment leads to better visual understanding. • Fine-grained inter-data similarity can serve as soft targets for cross-modal alignment.

In this paper, we design and evaluate a new substructure-aware Graph Representation Learning (GRL) approach. GRL aims to map graph structure information into low-dimensional representations. While extensive efforts have been made for modeling global and/or local structure information, GRL can be improved by substructure information. Some recent studies exploit adversarial learning to incorporate substructure awareness, but hindered by unstable convergence. This study will address the major research question: is there a better way to integrate substructure awareness into GRL? As subsets of the graph structure, interested substructures (i.e., subgraph) are unique and representative for differentiating graphs, leading to the high correlation between the representation of the graph-level structure and substructures. Since mutual information (MI) is to evaluate the mutual dependence between two variables, we develop a MI inducted substructure-aware GRL method. We decompose the GRL pipeline into two stages: (1) node-level, where we introduce to maximize MI between the original and learned representation by the intuition that the original and learned representation should be highly correlated; (2) graph-level, where we preserve substructures by maximizing MI between the graph-level structure and substructure representation. Finally, we present extensive experimental results to demonstrate the improved performances of our method with real-world data.

Recent advances in vision-language models have resulted in improved accuracy in visual question answering (VQA) tasks. However, their robustness remains limited when faced with out-of-distribution data containing unanswerable questions. In this study, we first construct a simple randomised VQA dataset, incorporating unanswerable questions from the VQA v2 dataset, to evaluate the robustness of a state-of-the-art VQA model. Our findings reveal that the model struggles to predict the “unknown” answer or provides inaccurate responses with high confidence scores for irrelevant questions. To address this issue without retraining the large backbone models, we propose Cross Modal Augmentation (CMA), a model-agnostic, test-time-only, multi-modal semantic augmentation technique. CMA generates multiple semantically-consistent but heterogeneous instances from the visual and textual inputs, which are then fed to the model, and the predictions are combined to achieve a more robust output. We demonstrate that implementing CMA enables the VQA model to provide more reliable answers in scenarios involving unanswerable questions, and show that the approach is generalisable across different categories of pre-trained vision language models.

Vision-Language (VL) models with the Two-Tower architecture have dominated visual-language representation learning in recent years. Current VL models either use lightweight uni-modal encoders and learn to extract, align and fuse both modalities simultaneously in a deep cross-modal encoder, or feed the last-layer uni-modal representations from the deep pre-trained uni-modal encoders into the top cross-modal encoder. Both approaches potentially restrict vision-language representation learning and limit model performance. In this paper, we propose BridgeTower, which introduces multiple bridge layers that build a connection between the top layers of uni-modal encoders and each layer of the cross-modal encoder. This enables effective bottom-up cross-modal alignment and fusion between visual and textual representations of different semantic levels of pre-trained uni-modal encoders in the cross-modal encoder. Pre-trained with only 4M images, BridgeTower achieves state-of-the-art performance on various downstream vision-language tasks. In particular, on the VQAv2 test-std set, BridgeTower achieves an accuracy of 78.73%, outperforming the previous state-of-the-art model METER by 1.09% with the same pre-training data and almost negligible additional parameters and computational costs. Notably, when further scaling the model, BridgeTower achieves an accuracy of 81.15%, surpassing models that are pre-trained on orders-of-magnitude larger datasets. Code and checkpoints are available at https://github.com/microsoft/BridgeTower.

Learning discriminative image representations plays a vital role in long-tailed image classification because it can ease the classifier learning in imbalanced cases. Given the promising performance contrastive learning has shown recently in representation learning, in this work, we explore effective supervised contrastive learning strategies and tailor them to learn better image representations from imbalanced data in order to boost the classification accuracy thereon. Specifically, we propose a novel hybrid network structure being composed of a supervised contrastive loss to learn image representations and a cross-entropy loss to learn classifiers, where the learning is progressively transited from feature learning to the classifier learning to embody the idea that better features make better classifiers. We explore two variants of contrastive loss for feature learning, which vary in the forms but share a common idea of pulling the samples from the same class together in the normalized embedding space and pushing the samples from different classes apart. One of them is the recently proposed supervised contrastive (SC) loss, which is designed on top of the state-of-the-art unsupervised contrastive loss by incorporating positive samples from the same class. The other is a prototypical supervised contrastive (PSC) learning strategy which addresses the intensive memory consumption in standard SC loss and thus shows more promise under limited memory budget. Extensive experiments on three long-tailed classification datasets demonstrate the advantage of the proposed contrastive learning based hybrid networks in long-tailed classification.

It is essential to delve into the strategy of multimodal model pre-training, which is an obvious impact on downstream tasks. Currently, clustering learning has achieved noteworthy benefits in multiple methods. However, due to the availability of open image-text pairs, it is challenging for multimodal with clustering learning. In this paper, we propose an approach that utilizes clustering swap prediction strategy to learn image-text clustering embedding space by interaction prediction between image and text features. Unlike existing models with clustering learning, our method (Clus) allows for an open number of clusters for web-scale alt-text data. Furthermore, in order to train the image and text encoders efficiently, we introduce distillation learning approach and evaluate the performance of the image-encoder in downstream visual tasks. In addition, Clus is pre-trained end-to-end by using large-scale image-text pairs. Specifically, both text and image serve as ground truth for swap prediction, enabling effective representation learning. Concurrently, extensive experiments demonstrate that Clus achieves state-of-the-art performance on multiple downstream fine-tuning and zero-shot tasks (i.e., Image-Text Retrieval, VQA, NLVR2, Image Captioning, Object Detection, and Semantic Segmentation).

Multimodal representations play an important role in multimodal learning tasks, including cross-modal retrieval and intra-modal clustering. However, existing multimodal representation learning approaches focus on building one common space by aligning different modalities and ignore the complementary information across the modalities, such as the intra-modal local structures. In other words, they only focus on the object-level alignment and ignore structure-level alignment. To tackle the problem, we propose a novel symmetric multimodal representation learning framework by transferring local structures across different modalities, namely MTLS. A customized soft metric learning strategy and an iterative parameter learning process are designed to symmetrically transfer local structures and enhance the cluster structures in intra-modal representations. The bidirectional retrieval loss based on multi-layer neural networks is utilized to align two modalities. MTLS is instantiated with image and text data and shows its superior performance on image-text retrieval and image clustering. MTLS outperforms the state-of-the-art multimodal learning methods by up to 32% in terms of R@1 on text-image retrieval and 16.4% in terms of AMI onclustering.

Autoregressive Predictive Coding (APC), as a self-supervised objective, has enjoyed success in learning representations from large amounts of unlabeled data, and the learned representations are rich for many downstream tasks.However, the connection between low self-supervised loss and strong performance in downstream tasks remains unclear.In this work, we propose Vector-Quantized Autoregressive Predictive Coding (VQ-APC), a novel model that produces quantized representations, allowing us to explicitly control the amount of information encoded in the representations.By studying a sequence of increasingly limited models, we reveal the constituents of the learned representations.In particular, we confirm the presence of information with probing tasks, while showing the absence of information with mutual information, uncovering the model's preference in preserving speech information as its capacity becomes constrained.We find that there exists a point where phonetic and speaker information are amplified to maximize a selfsupervised objective.As a byproduct, the learned codes for a particular model capacity correspond well to English phones.

This monograph surveys vision-language pre-training (VLP) methods for multimodal intelligence that have been developed in the last few years. We group these approaches into three categories: (i) VLP for image-text tasks, such as image captioning, image-text retrieval, visual question answering, and visual grounding; (ii) VLP for core computer vision tasks, such as (open-set) image classification, object detection, and segmentation; and (iii) VLP for video-text tasks, such as video captioning, video-text retrieval, and video question answering. For each category, we present a comprehensive review of state-of-the-art methods, and discuss the progress that has been made and challenges still being faced, using specific systems and models as case studies. In addition, for each category, we discuss advanced topics being actively explored in the research community, such as big foundation models, unified modeling, in-context few-shot learning, knowledge, robustness, and computer vision in the wild, to name a few.

Lying in the cross-section of computer vision and natural language processing, vision language models are capable of processing images and text at once. These models are helpful in various tasks: text generation from image and vice versa, image-text retrieval, or visual navigation. Besides building a model trained on a dataset for a task, people also study general-purpose models to utilize many datasets for multitasks. Their two primary applications are image captioning and visual question answering. For English, large datasets and foundation models are already abundant. However, for Vietnamese, they are still limited. To expand the language range, this work proposes a pretrained general-purpose image-text model named VisualRoBERTa. A dataset of 600k images with captions (translated MS COCO 2017 from English to Vietnamese) is introduced to pretrain VisualRoBERTa. The model’s architecture is built using Convolutional Neural Network and Transformer blocks. Fine-tuning VisualRoBERTa shows promising results on the ViVQA dataset with 34.49% accuracy, 0.4173 BLEU 4, and 0.4390 RougeL (in visual question answering task), and best outcomes on the sViIC dataset with 0.6685 BLEU 4, 0.6320 RougeL (in image captioning task).

MotivationProtein representation learning methods have shown great potential to many downstream tasks in biological applications. A few recent studies have demonstrated that the self-supervised learning is a promising solution to addressing insufficient labels of proteins, which is a major obstacle to effective protein representation learning. However, existing protein representation learning is usually pretrained on protein sequences without considering the important protein structural information.ResultsIn this work, we propose a novel structure-aware protein self-supervised learning method to effectively capture structural information of proteins. In particular, a graph neural network model is pretrained to preserve the protein structural information with self-supervised tasks from a pairwise residue distance perspective and a dihedral angle perspective, respectively. Furthermore, we propose to leverage the available protein language model pretrained on protein sequences to enhance the self-supervised learning. Specifically, we identify the relation between the sequential information in the protein language model and the structural information in the specially designed graph neural network model via a novel pseudo bi-level optimization scheme. We conduct experiments on three downstream tasks: the binary classification into membrane/non-membrane proteins, the location classification into 10 cellular compartments, and the enzyme-catalyzed reaction classification into 384 EC numbers, and these experiments verify the effectiveness of our proposed method.Availability and implementationThe Alphafold2 database is available in https://alphafold.ebi.ac.uk/. The PDB files are available in https://www.rcsb.org/. The downstream tasks are available in https://github.com/phermosilla/IEConv\\_proteins/tree/master/Datasets. The code of the proposed method is available in https://github.com/GGchen1997/STEPS_Bioinformatics.