Date of Award


Degree Type


Degree Name

Doctor of Philosophy (PhD)


Electrical Engineering and Computer Science


Qinru Qiu

Second Advisor

Lixin Shen


Visual-semantic learning is an attractive and challenging research direction aiming to understand complex semantics of heterogeneous data from two domains, i.e., visual signals (i.e., images and videos) and natural language (i.e., captions and questions). It requires memorizing the rich information in a single modality and a joint comprehension of multiple modalities. Artificial intelligence (AI) systems with human-level intelligence are claimed to learn like humans, such as efficiently leveraging brain memory for better comprehension, rationally incorporating common-sense knowledge into reasoning, quickly gaining in-depth understanding given a few samples, and analyzing relationships among abundant and informative events. However, these intelligence capacities are effortless for humans but challenging for machines. To bridge the discrepancy between human-level intelligence and present-day visual-semantic learning, we start from its basic understanding ability by studying the visual question answering (e.g., Image-QA and Video-QA) tasks from the perspectives of memory augmentation and common-sense knowledge incorporation. Furthermore, we stretch it to a more challenging situation with limited and partially unlabeled training data (i.e., Few-shot Visual-Semantic Learning) to imitate the fast learning ability of humans. Finally, to further enhance visual-semantic performance in natural videos with numerous spatio-temporal dynamics, we investigate exploiting event-correlated information for a comprehensive understanding of cross-modal semantics.

To study the essential visual-semantic understanding ability of the human brain with memory, we first propose a novel Memory Augmented Deep Recurrent Neural Network (i.e., MA-DRNN) model for Video-QA, which features a new method for encoding videos and questions, and memory augmentation using the emerging Differentiable Neural Computer (i.e., DNC). Specifically, we encode semantic (i.e., questions) information before visual (i.e., videos) information, which leads to better visual-semantic representations. Moreover, we leverage Differentiable Neural Computer (with external memory) to store and retrieve valuable information in questions and videos and model the long-term visual-semantic dependency.

In addition to basic understanding, to tackle visual-semantic reasoning that requires external knowledge beyond visible contents (e.g., KB-Image-QA), we propose a novel framework that endows the model with capabilities of answering more general questions and achieves better exploitation of external knowledge through generating Multiple Clues for Reasoning with Memory Neural Networks (i.e., MCR-MemNN). Specifically, a well-defined detector is adopted to predict image-question-related relation phrases, each delivering two complementary clues to retrieve the supporting facts from an external knowledge base (i.e., KB). These facts are encoded into a continuous embedding space using a content-addressable memory. Afterward, mutual interactions between visual-semantic representation and the supporting facts stored in memory are captured to distill the most relevant information in three modalities (i.e., image, question, and KB). Finally, the optimal answer is predicted by choosing the supporting fact with the highest score.

Furthermore, to enable a fast, in-depth understanding given a small number of samples, especially with heterogeneity in the multi-modal scenarios such as image question answering (i.e., Image-QA) and image captioning (i.e., IC), we study the few-shot visual-semantic learning and present the Hierarchical Graph ATtention Network (i.e., HGAT). This two-stage network models the intra- and inter-modal relationships with limited image-text samples. The main contributions of HGAT can be summarized as follows: 1) it sheds light on tackling few-shot multi-modal learning problems, which focuses primarily, but not exclusively, on visual and semantic modalities, through better exploitation of the intra-relationship of each modality and an attention-based co-learning framework between modalities using a hierarchical graph-based architecture; 2) it achieves superior performance on both visual question answering and image captioning in the few-shot setting; 3) it can be easily extended to the semi-supervised setting where image-text samples are partially unlabeled.

Although various attention mechanisms have been utilized to manage contextualized representations by modeling intra- and inter-modal relationships of the two modalities, one limitation of the predominant visual-semantic methods is the lack of reasoning with event correlation, sensing, and analyzing relationships among abundant and informative events contained in the video. To this end, we introduce the dense caption modality as a new auxiliary and distill event-correlated information to infer the correct answer. We propose a novel end-to-end trainable model, Event-Correlated Graph Neural Networks (EC-GNNs), to perform cross-modal reasoning over information from the three modalities (i.e., caption, video, and question). Besides exploiting a new modality, we employ cross-modal reasoning modules to explicitly model inter-modal relationships and aggregate relevant information across different modalities. We propose a question-guided self-adaptive multi-modal fusion module to collect the question-oriented and event-correlated evidence through multi-step reasoning.

To evaluate our proposed models, we conduct extensive experiments on VTW, MSVD-QA, and TGIF-QA datasets for Video-QA task, Toronto COCO-QA, Visual Genome-QA datasets for few-shot Image-QA task, COCO-FITB dataset for few-shot IC task, and FVQA, Visual7W + ConceptNet datasets for KB-Image-QA task. The experimental results justify these models’ effectiveness and superiority over baseline methods.


Open Access