Bilinear Graph Networks for Visual Question Answering

Bridging the Cross-Modality Semantic Gap in Visual Question Answering

The objective of visual question answering (VQA) is to adequately comprehend a question and identify relevant contents in an image that can provide an answer. Existing approaches in VQA often combine visual and question features directly to create a unified cross-modality representation for answer inference. However, this kind of approach fails to bridge the semantic gap between visual and text modalities, resulting in a lack of alignment in cross-modality semantics and the inability to match key visual content accurately. In this article, we propose a model called the caption bridge-based cross-modality alignment and contrastive learning model (CBAC) to address the issue. The CBAC model aims to reduce the semantic gap between different modalities. It consists of a caption-based cross-modality alignment module and a visual-caption (V-C) contrastive learning module. By utilizing an auxiliary caption that shares the same modality as the question and has closer semantic associations with the visual, we are able to effectively reduce the semantic gap by separately matching the caption with both the question and the visual to generate pre-alignment features for each, which are then used in the subsequent fusion process. We also leverage the fact that V-C pairs exhibit stronger semantic connections compared to question-visual (Q-V) pairs to employ a contrastive learning mechanism on visual and caption pairs to further enhance the semantic alignment capabilities of single-modality encoders. Extensive experiments conducted on three benchmark datasets demonstrate that the proposed model outperforms previous state-of-the-art VQA models. Additionally, ablation experiments confirm the effectiveness of each module in our model. Furthermore, we conduct a qualitative analysis by visualizing the attention matrices to assess the reasoning reliability of the proposed model.

Latent Attention Network With Position Perception for Visual Question Answering

For exploring the complex relative position relationships among multiobject with multiple position prepositions in the question, we propose a novel latent attention (LA) network for visual question answering (VQA), in which LA with position perception is extracted by a novel LA generation module (LAGM) and encoded along with absolute and relative position relations by our proposed position-aware module (PAM). The LAGM reconstructs original attention into LA by capturing the tendency of visual attention shifting according to the position prepositions in the question. The LA accurately captures the complex relative position features of multiple objects and helps the model locate the attention to the correct object or region. The PAM adopts latent state and relative position relations to enhance the capability of comprehending the multiobject correlations. In addition, we also propose a novel gated counting module (GCM) to strengthen the sensitivity of quantitative knowledge for effectively improving the performance of counting questions. Extensive experiments demonstrate that our proposed method achieves excellent performance on VQA and outperforms state-of-the-art methods on the widely used datasets VQA v2 and VQA v1.

Achieving Sharp Upper Bounds on the Expressive Power of Neural Networks via Tropical Polynomials

The expressive power of neural networks describes the ability to represent or approximate complex functions. The number of linear regions is the standard and most natural measure of expressive power. However, a major challenge in utilizing the number of linear regions as a measure of expressive power is the exponential gap between the theoretical upper and lower bounds, which becomes more pronounced as the neural network capacity increases. In this article, we aim to derive a sharp upper bound on piecewise linear neural networks (PLNNs) to bridge this gap. Specifically, we first establish the relationship between tropical polynomials and PLNNs. In the unexpanded tropical polynomials form, we make the proposition that hyperplanes are not all in the general positions, thereby reducing the number of intersecting hyperplanes. We propose a rank-based approach and present the empirical analysis that this approach outperforms previous Zaslavsky's theorem-based methods. In the expanded tropical polynomials form, accounting for limitations in weight initialization and model computational precision, we raise the concept that the values range of each term is bounded. We propose a precision-based approach that transforms the approximate exponential growth of the number of linear regions into polynomial growth with width, which is effective at larger layer widths. Finally, we compare the number of linear regions that can be represented by each hidden layer in both forms and derive a sharp upper bound for PLNNs. Empirical analysis and experimental results provide compelling evidence for the efficacy and feasibility of this sharp upper bound on both simulated experiments and real datasets.

MLW-BFECF: A Multi-Weighted Dynamic Cascade Forest Based on Bilinear Feature Extraction for Predicting the Stage of Kidney Renal Clear Cell Carcinoma on Multi-Modal Gene Data

The stage prediction of kidney renal clear cell carcinoma (KIRC) is important for the diagnosis, personalized treatment, and prognosis of patients. Many prediction methods have been proposed, but most of them are based on unimodal gene data, and their accuracy is difficult to further improve. Therefore, we propose a novel multi-weighted dynamic cascade forest based on the bilinear feature extraction (MLW-BFECF) model for stage prediction of KIRC using multimodal gene data (RNA-seq, CNA, and methylation). The proposed model utilizes a dynamic cascade framework with shuffle layers to prevent early degradation of the model. In each cascade layer, a voting technique based on three gene selection algorithms is first employed to effectively retain gene features more relevant to KIRC and eliminate redundant information in gene features. Then, two new bilinear models based on the gated attention mechanism are proposed to better extract new intra-modal and inter-modal gene features; Finally, based on the idea of the bagging, a multi-weighted ensemble forest classifiers module is proposed to extract and fuse probabilistic features of the three-modal gene data. A series of experiments demonstrate that the MLW-BFECF model based on the three-modal KIRC dataset achieves the highest prediction performance with an accuracy of 88.9 %.

To Boost Zero-Shot Generalization for Embodied Reasoning With Vision-Language Pre-Training

Recently, there exists an increased research interest in embodied artificial intelligence (EAI), which involves an agent learning to perform a specific task when dynamically interacting with the surrounding 3D environment. There into, a new challenge is that many unseen objects may appear due to the increased number of object categories in 3D scenes. It makes developing models with strong zero-shot generalization ability to new objects necessary. Existing work tries to achieve this goal by providing embodied agents with massive high-quality human annotations closely related to the task to be learned, while it is too costly in practice. Inspired by recent advances in pre-trained models in 2D visual tasks, we attempt to boost zero-shot generalization for embodied reasoning with vision-language pre-training that can encode common sense as general prior knowledge. To further improve its performance on a specific task, we rectify the pre-trained representation through masked scene graph modeling (MSGM) in a self-supervised manner, where the task-specific knowledge is learned from iterative message passing. Our method can improve a variety of representative embodied reasoning tasks by a large margin (e.g., over 5.0% w.r.t. answer accuracy on MP3D-EQA dataset that consists of many real-world scenes with a large number of new objects during testing), and achieve the new state-of-the-art performance.

Motif Graph Neural Network

Graphs can model complicated interactions between entities, which naturally emerge in many important applications. These applications can often be cast into standard graph learning tasks, in which a crucial step is to learn low-dimensional graph representations. Graph neural networks (GNNs) are currently the most popular model in graph embedding approaches. However, standard GNNs in the neighborhood aggregation paradigm suffer from limited discriminative power in distinguishing high-order graph structures as opposed to low-order structures. To capture high-order structures, researchers have resorted to motifs and developed motif-based GNNs. However, the existing motif-based GNNs still often suffer from less discriminative power on high-order structures. To overcome the above limitations, we propose motif GNN (MGNN), a novel framework to better capture high-order structures, hinging on our proposed motif redundancy minimization operator and injective motif combination. First, MGNN produces a set of node representations with respect to each motif. The next phase is our proposed redundancy minimization among motifs which compares the motifs with each other and distills the features unique to each motif. Finally, MGNN performs the updating of node representations by combining multiple representations from different motifs. In particular, to enhance the discriminative power, MGNN uses an injective function to combine the representations with respect to different motifs. We further show that our proposed architecture increases the expressive power of GNNs with a theoretical analysis. We demonstrate that MGNN outperforms state-of-the-art methods on seven public benchmarks on both the node classification and graph classification tasks.

Relation-Aware Heterogeneous Graph Network for Learning Intermodal Semantics in Textbook Question Answering

Textbook question answering (TQA) task aims to infer answers for given questions from a multimodal context, including text and diagrams. The existing studies have aggregated intramodal semantics extracted from a single modality but have yet to capture the intermodal semantics between different modalities. A major challenge in learning intermodal semantics is maintaining lossless intramodal semantics while bridging the gap of semantics caused by heterogeneity. In this article, we propose an intermodal relation-aware heterogeneous graph network (IMR-HGN) to extract the intermodal semantics for TQA, which aggregates different modalities while learning features rather than representing them independently. First, we design a multidomain consistent representation (MDCR) to eliminate semantic gaps by capturing intermodal features while maintaining lossless intramodal semantics in multidomains. Furthermore, we present neighbor-based relation inpainting (NRI) to reduce semantic ambiguity via repairing fuzzy relations with correlations of relations. Finally, we propose hierarchical multisemantics aggregation (HMSA) to guarantee the completeness of semantics by aggregating features of nodes and relations with a reconstruction network (RN). Experimental results show that IMR-HGN could extract the intermodal semantics of answers, achieving a 2.16% improvement on the validation set of the TQA dataset and a 3.04% increase on the test set of the AI2D dataset.

Multi-Modal Alignment of Visual Question Answering Based on Multi-Hop Attention Mechanism

The alignment of information between the image and the question is of great significance in the visual question answering (VQA) task. Self-attention is commonly used to generate attention weights between image and question. These attention weights can align two modalities. Through the attention weight, the model can select the relevant area of the image to align with the question. However, when using the self-attention mechanism, the attention weight between two objects is only determined by the representation of these two objects. It ignores the influence of other objects around these two objects. This contribution proposes a novel multi-hop attention alignment method that enriches surrounding information when using self-attention to align two modalities. Simultaneously, in order to utilize position information in alignment, we also propose a position embedding mechanism. The position embedding mechanism extracts the position information of each object and implements the position embedding mechanism to align the question word with the correct position in the image. According to the experiment on the VQA2.0 dataset, our model achieves validation accuracy of 65.77%, outperforming several state-of-the-art methods. The experimental result shows that our proposed methods have better performance and effectiveness.