Shu Zhao

Huanjing Zhao, Beining Yang, Yukuo Cen et al. · tsinghua

The text-attributed graph (TAG) is one kind of important real-world graph-structured data with each node associated with raw texts. For TAGs, traditional few-shot node classification methods directly conduct training on the pre-processed node features and do not consider the raw texts. The performance is highly dependent on the choice of the feature pre-processing method. In this paper, we propose P2TAG, a framework designed for few-shot node classification on TAGs with graph pre-training and prompting. P2TAG first pre-trains the language model (LM) and graph neural network (GNN) on TAGs with self-supervised loss. To fully utilize the ability of language models, we adapt the masked language modeling objective for our framework. The pre-trained model is then used for the few-shot node classification with a mixed prompt method, which simultaneously considers both text and graph information. We conduct experiments on six real-world TAGs, including paper citation networks and product co-purchasing networks. Experimental results demonstrate that our proposed framework outperforms existing graph few-shot learning methods on these datasets with +18.98% ~ +35.98% improvements.

Wenjie Zheng, Wenxue Wang, Shu Zhao et al.

Knowledge graph embeddings (KGE) have been validated as powerful methods for inferring missing links in knowledge graphs (KGs) that they typically map entities into Euclidean space and treat relations as transformations of entities. Recently, some Euclidean KGE methods have been enhanced to model semantic hierarchies commonly found in KGs, improving the performance of link prediction. To embed hierarchical data, hyperbolic space has emerged as a promising alternative to traditional Euclidean space, offering high fidelity and lower memory consumption. Unlike Euclidean, hyperbolic space provides countless curvatures to choose from. However, it is difficult for existing hyperbolic KGE methods to obtain the optimal curvature settings manually, thereby limiting their ability to effectively model semantic hierarchies. To address this limitation, we propose a novel KGE model called $\textbf{Hyp}$erbolic $\textbf{H}$ierarchical $\textbf{KGE}$ (HypHKGE). This model introduces attention-based learnable curvatures for hyperbolic space, which helps preserve rich semantic hierarchies. Furthermore, to utilize the preserved hierarchies for inferring missing links, we define hyperbolic hierarchical transformations based on the theory of hyperbolic geometry, including both inter-level and intra-level modeling. Experiments demonstrate the effectiveness of the proposed HypHKGE model on the three benchmark datasets (WN18RR, FB15K-237, and YAGO3-10). The source code will be publicly released at https://github.com/wjzheng96/HypHKGE.

Shicheng Tan, Weng Lam Tam, Yuanchun Wang et al.

Currently, the reduction in the parameter scale of large-scale pre-trained language models (PLMs) through knowledge distillation has greatly facilitated their widespread deployment on various devices. However, the deployment of knowledge distillation systems faces great challenges in real-world industrial-strength applications, which require the use of complex distillation methods on even larger-scale PLMs (over 10B), limited by memory on GPUs and the switching of methods. To overcome these challenges, we propose GKD, a general knowledge distillation framework that supports distillation on larger-scale PLMs using various distillation methods. With GKD, developers can build larger distillation models on memory-limited GPUs and easily switch and combine different distillation methods within a single framework. Experimental results show that GKD can support the distillation of at least 100B-scale PLMs and 25 mainstream methods on 8 NVIDIA A100 (40GB) GPUs.

Zijian Zhu, Yichi Zhang, Hai Chen et al.

3D object detection is an essential perception task in autonomous driving to understand the environments. The Bird's-Eye-View (BEV) representations have significantly improved the performance of 3D detectors with camera inputs on popular benchmarks. However, there still lacks a systematic understanding of the robustness of these vision-dependent BEV models, which is closely related to the safety of autonomous driving systems. In this paper, we evaluate the natural and adversarial robustness of various representative models under extensive settings, to fully understand their behaviors influenced by explicit BEV features compared with those without BEV. In addition to the classic settings, we propose a 3D consistent patch attack by applying adversarial patches in the 3D space to guarantee the spatiotemporal consistency, which is more realistic for the scenario of autonomous driving. With substantial experiments, we draw several findings: 1) BEV models tend to be more stable than previous methods under different natural conditions and common corruptions due to the expressive spatial representations; 2) BEV models are more vulnerable to adversarial noises, mainly caused by the redundant BEV features; 3) Camera-LiDAR fusion models have superior performance under different settings with multi-modal inputs, but BEV fusion model is still vulnerable to adversarial noises of both point cloud and image. These findings alert the safety issue in the applications of BEV detectors and could facilitate the development of more robust models.

Yelin Chen, Fanjin Zhang, Suping Sun et al.

Understanding research papers remains challenging for foundation models due to specialized scientific discourse and complex figures and tables, yet existing benchmarks offer limited fine-grained evaluation at scale. To address this gap, we introduce RPC-Bench, a large-scale question-answering benchmark built from review-rebuttal exchanges of high-quality computer science papers, containing 15K human-verified QA pairs. We design a fine-grained taxonomy aligned with the scientific research flow to assess models' ability to understand and answer why, what, and how questions in scholarly contexts. We also define an elaborate LLM-human interaction annotation framework to support large-scale labeling and quality control. Following the LLM-as-a-Judge paradigm, we develop a scalable framework that evaluates models on correctness-completeness and conciseness, with high agreement to human judgment. Experiments reveal that even the strongest models (GPT-5) achieve only 68.2% correctness-completeness, dropping to 37.46% after conciseness adjustment, highlighting substantial gaps in precise academic paper understanding. Our code and data are available at https://rpc-bench.github.io/.

Shicheng Tan, Weng Lam Tam, Yuanchun Wang et al.

The large scale of pre-trained language models poses a challenge for their deployment on various devices, with a growing emphasis on methods to compress these models, particularly knowledge distillation. However, current knowledge distillation methods rely on the model's intermediate layer features and the golden labels (also called hard labels), which usually require aligned model architecture and enough labeled data respectively. Moreover, the parameters of vocabulary are usually neglected in existing methods. To address these problems, we propose a general language model distillation (GLMD) method that performs two-stage word prediction distillation and vocabulary compression, which is simple and surprisingly shows extremely strong performance. Specifically, GLMD supports more general application scenarios by eliminating the constraints of dimension and structure between models and the need for labeled datasets through the absence of intermediate layers and golden labels. Meanwhile, based on the long-tailed distribution of word frequencies in the data, GLMD designs a strategy of vocabulary compression through decreasing vocabulary size instead of dimensionality. Experimental results show that our method outperforms 25 state-of-the-art methods on the SuperGLUE benchmark, achieving an average score that surpasses the best method by 3%.

Shu Zhao, Almıla Akdağ Salah, Albert Ali Salah

The way the human body is depicted in classical and modern paintings is relevant for art historical analyses. Each artist has certain themes and concerns, resulting in different poses being used more heavily than others. In this paper, we propose a computer vision pipeline to analyse human pose and representations in paintings, which can be used for specific artists or periods. Specifically, we combine two pose estimation approaches (OpenPose and DensePose, respectively) and introduce methods to deal with occlusion and perspective issues. For normalisation, we map the detected poses and contours to Leonardo da Vinci's Vitruvian Man, the classical depiction of body proportions. We propose a visualisation approach for illustrating the articulation of joints in a set of paintings. Combined with a hierarchical clustering of poses, our approach reveals common and uncommon poses used by artists. Our approach improves over purely skeleton based analyses of human body in paintings.

Le Zhang, Yong Gu, Xin Zhao et al.

ATARI is a suite of video games used by reinforcement learning (RL) researchers to test the effectiveness of the learning algorithm. Receiving only the raw pixels and the game score, the agent learns to develop sophisticated strategies, even to the comparable level of a professional human games tester. Ideally, we also want an agent requiring very few interactions with the environment. Previous competitive model-free algorithms for the task use the valued-based Rainbow algorithm without any policy head. In this paper, we change it by proposing a practical discrete variant of the soft actor-critic (SAC) algorithm. The new variant enables off-policy learning using policy heads for discrete domains. By incorporating it into the advanced Rainbow variant, i.e., the ``bigger, better, faster'' (BBF), the resulting SAC-BBF improves the previous state-of-the-art interquartile mean (IQM) from 1.045 to 1.088, and it achieves these results using only replay ratio (RR) 2. By using lower RR 2, the training time of SAC-BBF is strictly one-third of the time required for BBF to achieve an IQM of 1.045 using RR 8. As a value of IQM greater than one indicates super-human performance, SAC-BBF is also the only model-free algorithm with a super-human level using only RR 2. The code is publicly available on GitHub at https://github.com/lezhang-thu/bigger-better-faster-SAC.

Muyu Pan, Shu Zhao, Nan Zhang et al.

This paper investigates large language model (LLM) abstention learning, specifically using ternary reward, which incentivize truthfulness in large language models. This paper extends that idea by moving from a ternary reward to a Trajectory-Informed advantage reweighting, dynamically re-weights the abstention reward during Group Relative Policy Optimization (GRPO) training. The objective of this work focuses on abstention learning instead of improving truthfulness, serving as an exploration into hallucination reduction. The novelty of this paper lies in methodological innovation, advantage re-weighting, and benchmark selection. Leveraging GRPO's multiple trajectories as a natural abstention signal, this method uses a reward signal to explore knowledge boundaries and encourage consistency. By demonstrating that trajectories can be used as a confidence indicator of the policy relative to the query, they are then used to dynamically calculate the abstention advantage. AbstentionBench is used as the evaluation benchmark, as this work aims to contribute to the field of abstention learning. All datasets on the benchmark were tested against this method and various baselines. Empirical results demonstrate that TIAR achieves state-of-the-art abstention F1 scores across five of six evaluation categories, outperforming the static ternary baseline on 17 of 31 benchmark datasets while fully preserving baseline accuracy.

Xiaokang Zhang, Sijia Luo, Bohan Zhang et al. · tsinghua

We introduce TableLLM, a robust large language model (LLM) with 8 billion parameters, purpose-built for proficiently handling tabular data manipulation tasks, whether they are embedded within documents or spreadsheets, catering to real-world office scenarios. We propose a distant supervision method for training, which comprises a reasoning process extension strategy, aiding in training LLMs to understand reasoning patterns more effectively as well as a cross-way validation strategy, ensuring the quality of the automatically generated data. To evaluate the performance of TableLLM, we have crafted benchmarks tailored to address both document and spreadsheet formats as well as constructed a well-organized evaluation pipeline capable of handling both scenarios. Thorough evaluations underscore the advantages of TableLLM when compared to various existing general-purpose and tabular data-focused LLMs. We have publicly released the model checkpoint, source code, benchmarks, and a web application for user interaction. Our codes and data are publicly available at https://github.com/TableLLM/TableLLM.

Shu Zhao, Huijuan Xu

Composed image retrieval which combines a reference image and a text modifier to identify the desired target image is a challenging task, and requires the model to comprehend both vision and language modalities and their interactions. Existing approaches focus on holistic multi-modal interaction modeling, and ignore the composed and complimentary property between the reference image and text modifier. In order to better utilize the complementarity of multi-modal inputs for effective information fusion and retrieval, we move the multi-modal understanding to fine-granularity at concept-level, and learn the multi-modal concept alignment to identify the visual location in reference or target images corresponding to text modifier. Toward the end, we propose a NEUral COncept REasoning (NEUCORE) model which incorporates multi-modal concept alignment and progressive multimodal fusion over aligned concepts. Specifically, considering that text modifier may refer to semantic concepts not existing in the reference image and requiring to be added into the target image, we learn the multi-modal concept alignment between the text modifier and the concatenation of reference and target images, under multiple-instance learning framework with image and sentence level weak supervision. Furthermore, based on aligned concepts, to form discriminative fusion features of the input modalities for accurate target image retrieval, we propose a progressive fusion strategy with unified execution architecture instantiated by the attended language semantic concepts. Our proposed approach is evaluated on three datasets and achieves state-of-the-art results.

Xin Jin, Shu Zhao, Le Zhang et al.

In recent years, image generation has made great strides in improving the quality of images, producing high-fidelity ones. Also, quite recently, there are architecture designs, which enable GAN to unsupervisedly learn the semantic attributes represented in different layers. However, there is still a lack of research on generating face images more consistent with human aesthetics. Based on EigenGAN [He et al., ICCV 2021], we build the techniques of reinforcement learning into the generator of EigenGAN. The agent tries to figure out how to alter the semantic attributes of the generated human faces towards more preferable ones. To accomplish this, we trained an aesthetics scoring model that can conduct facial beauty prediction. We also can utilize this scoring model to analyze the correlation between face attributes and aesthetics scores. Empirically, using off-the-shelf techniques from reinforcement learning would not work well. So instead, we present a new variant incorporating the ingredients emerging in the reinforcement learning communities in recent years. Compared to the original generated images, the adjusted ones show clear distinctions concerning various attributes. Experimental results using the MindSpore, show the effectiveness of the proposed method. Altered facial images are commonly more attractive, with significantly improved aesthetic levels.

Shu Zhao, Nilesh Ahuja, Tan Yu et al.

Pre-trained vision language models have shown remarkable performance on visual recognition tasks, but they typically assume the availability of complete multimodal inputs during both training and inference. In real-world scenarios, however, modalities may be missing due to privacy constraints, collection difficulties, or resource limitations. While previous approaches have addressed this challenge using prompt learning techniques, they fail to capture the cross-modal relationships necessary for effective multimodal visual recognition and suffer from inevitable computational overhead. In this paper, we introduce MoRA, a parameter-efficient fine-tuning method that explicitly models cross-modal interactions while maintaining modality-specific adaptations. MoRA introduces modality-common parameters between text and vision encoders, enabling bidirectional knowledge transfer. Additionally, combined with the modality-specific parameters, MoRA allows the backbone model to maintain inter-modality interaction and enable intra-modality flexibility. Extensive experiments on standard benchmarks demonstrate that MoRA achieves an average performance improvement in missing-modality scenarios by 5.24% and uses only 25.90% of the inference time compared to the SOTA method while requiring only 0.11% of trainable parameters compared to full fine-tuning.

Shu Zhao, Xiaohan Zou, Tan Yu et al.

Pre-trained large multi-modal models (LMMs) exploit fine-tuning to adapt diverse user applications. Nevertheless, fine-tuning may face challenges due to deactivated sensors (e.g., cameras turned off for privacy or technical issues), yielding modality-incomplete data and leading to inconsistency in training data and the data for inference. Additionally, continuous training leads to catastrophic forgetting, diluting the knowledge in pre-trained LMMs. To overcome these challenges, we introduce a novel task, Continual Missing Modality Learning (CMML), to investigate how models can generalize when data of certain modalities is missing during continual fine-tuning. Our preliminary benchmarks reveal that existing methods suffer from a significant performance drop in CMML, even with the aid of advanced continual learning techniques. Therefore, we devise a framework termed Reconstruct before Query (RebQ). It decomposes prompts into modality-specific ones and breaks them into components stored in pools accessible via a key-query mechanism, which facilitates ParameterEfficient Fine-Tuning and enhances knowledge transferability for subsequent tasks. Meanwhile, our RebQ leverages extensive multi-modal knowledge from pre-trained LMMs to reconstruct the data of missing modality. Comprehensive experiments demonstrate that RebQ effectively reconstructs the missing modality information and retains pre-trained knowledge. Specifically, compared with the baseline, RebQ improves average precision from 20.00 to 50.92 and decreases average forgetting from 75.95 to 8.56. Code and datasets are available on https://github.com/Tree-Shu-Zhao/RebQ.pytorch

Xin Jin, Wu Zhou, Xinghui Zhou et al.

Aesthetic assessment of images can be categorized into two main forms: numerical assessment and language assessment. Aesthetics caption of photographs is the only task of aesthetic language assessment that has been addressed. In this paper, we propose a new task of aesthetic language assessment: aesthetic visual question and answering (AVQA) of images. If we give a question of images aesthetics, model can predict the answer. We use images from \textit{www.flickr.com}. The objective QA pairs are generated by the proposed aesthetic attributes analysis algorithms. Moreover, we introduce subjective QA pairs that are converted from aesthetic numerical labels and sentiment analysis from large-scale pre-train models. We build the first aesthetic visual question answering dataset, AesVQA, that contains 72,168 high-quality images and 324,756 pairs of aesthetic questions. Two methods for adjusting the data distribution have been proposed and proved to improve the accuracy of existing models. This is the first work that both addresses the task of aesthetic VQA and introduces subjectiveness into VQA tasks. The experimental results reveal that our methods outperform other VQA models on this new task.

Xiaoshuai Hao, Yunfeng Diao, Mengchuan Wei et al.

Map construction task plays a vital role in providing precise and comprehensive static environmental information essential for autonomous driving systems. Primary sensors include cameras and LiDAR, with configurations varying between camera-only, LiDAR-only, or camera-LiDAR fusion, based on cost-performance considerations. While fusion-based methods typically perform best, existing approaches often neglect modality interaction and rely on simple fusion strategies, which suffer from the problems of misalignment and information loss. To address these issues, we propose MapFusion, a novel multi-modal Bird's-Eye View (BEV) feature fusion method for map construction. Specifically, to solve the semantic misalignment problem between camera and LiDAR BEV features, we introduce the Cross-modal Interaction Transform (CIT) module, enabling interaction between two BEV feature spaces and enhancing feature representation through a self-attention mechanism. Additionally, we propose an effective Dual Dynamic Fusion (DDF) module to adaptively select valuable information from different modalities, which can take full advantage of the inherent information between different modalities. Moreover, MapFusion is designed to be simple and plug-and-play, easily integrated into existing pipelines. We evaluate MapFusion on two map construction tasks, including High-definition (HD) map and BEV map segmentation, to show its versatility and effectiveness. Compared with the state-of-the-art methods, MapFusion achieves 3.6% and 6.2% absolute improvements on the HD map construction and BEV map segmentation tasks on the nuScenes dataset, respectively, demonstrating the superiority of our approach.

Shu Zhao, Tan Yu, Anbang Xu et al.

Reasoning-augmented search agents such as Search-R1, trained via reinforcement learning with verifiable rewards (RLVR), demonstrate remarkable capabilities in multi-step information retrieval from external knowledge sources. These agents address the limitations of their parametric memory by dynamically gathering relevant facts to address complex reasoning tasks. However, existing approaches suffer from a fundamental architectural limitation: they process search queries strictly sequentially, even when handling inherently parallelizable and logically independent comparisons. This sequential bottleneck significantly constrains computational efficiency, particularly for queries that require multiple entity comparisons. To address this critical limitation, we propose ParallelSearch, a novel reinforcement learning framework that empowers large language models (LLMs) to recognize parallelizable query structures and execute multiple search operations concurrently. Our approach introduces dedicated reward functions that incentivize the identification of independent query components while preserving answer accuracy through jointly considering correctness, query decomposition quality, and parallel execution benefits. Comprehensive experiments demonstrate that ParallelSearch outperforms state-of-the-art baselines by an average performance gain of 2.9% across seven question-answering benchmarks. Notably, on parallelizable questions, our method achieves a 12.7% performance improvement while requiring only 69.6% of the LLM calls compared to sequential approaches.

Xiaoshuai Hao, Lingdong Kong, Rong Yin et al.

Robust high-definition (HD) map construction is vital for autonomous driving, yet existing methods often struggle with incomplete multi-view camera data. This paper presents SafeMap, a novel framework specifically designed to secure accuracy even when certain camera views are missing. SafeMap integrates two key components: the Gaussian-based Perspective View Reconstruction (G-PVR) module and the Distillation-based Bird's-Eye-View (BEV) Correction (D-BEVC) module. G-PVR leverages prior knowledge of view importance to dynamically prioritize the most informative regions based on the relationships among available camera views. Furthermore, D-BEVC utilizes panoramic BEV features to correct the BEV representations derived from incomplete observations. Together, these components facilitate the end-to-end map reconstruction and robust HD map generation. SafeMap is easy to implement and integrates seamlessly into existing systems, offering a plug-and-play solution for enhanced robustness. Experimental results demonstrate that SafeMap significantly outperforms previous methods in both complete and incomplete scenarios, highlighting its superior performance and reliability.

Shu Zhao, Tan Yu, Anbang Xu

Reasoning-augmented search agents, such as Search-R1, are trained to reason, search, and generate the final answer iteratively. Nevertheless, due to their limited capabilities in reasoning and search, their performance on multi-hop QA benchmarks remains far from satisfactory. To handle complex or compound queries, we train an LLM-based search agent with the native capability of query expansion through reinforcement learning. In each turn, our search agent proposes several query variants, which are searched simultaneously to cover more relevant information. Meanwhile, given limited post-training data and computing resources, it is very challenging for a search agent to master multiple tasks, including query generation, retrieved information understanding, and answer generation. Therefore, we propose incorporating a pre-trained squeezer model that helps the search agent understand the retrieved documents, allowing the search agent to focus on query generation for high retrieval recall. With the assistance of the squeezer model, we discover that even a small-scale 3B LLM can demonstrate a strong capability of query expansion and achieve state-of-the-art accuracy on the multi-hop QA benchmarks. To be specific, our experiments across seven question-answering benchmarks demonstrate that our method, named ExpandSearch, achieves an average improvement of 4.4% compared to state-of-the-art baselines, with strong gains on multi-hop reasoning tasks requiring diverse evidence aggregation.

Shu Zhao, Tan Yu, Xiaoshuai Hao et al.

Deep hashing has been widely used for large-scale approximate nearest neighbor search due to its storage and search efficiency. However, existing deep hashing methods predominantly rely on abundant training data, leaving the more challenging scenario of low-resource adaptation for deep hashing relatively underexplored. This setting involves adapting pre-trained models to downstream tasks with only an extremely small number of training samples available. Our preliminary benchmarks reveal that current methods suffer significant performance degradation due to the distribution shift caused by limited training samples. To address these challenges, we introduce Class-Calibration LoRA (CLoRA), a novel plug-and-play approach that dynamically constructs low-rank adaptation matrices by leveraging class-level textual knowledge embeddings. CLoRA effectively incorporates prior class knowledge as anchors, enabling parameter-efficient fine-tuning while maintaining the original data distribution. Furthermore, we propose Knowledge-Guided Discrete Optimization (KIDDO), a framework to utilize class knowledge to compensate for the scarcity of visual information and enhance the discriminability of hash codes. Extensive experiments demonstrate that our proposed method, Knowledge- Anchored Low-Resource Adaptation Hashing (KALAHash), significantly boosts retrieval performance and achieves a 4x data efficiency in low-resource scenarios.

Shu Zhao, Tianyi Shen, Nilesh Ahuja et al.

Multimodal Retrieval-Augmented Generation (MRAG) has emerged as a promising method to generate factual and up-to-date responses of Multimodal Large Language Models (MLLMs) by incorporating non-parametric knowledge from external knowledge bases. However, existing MRAG approaches suffer from static retrieval strategies, inflexible modality selection, and suboptimal utilization of retrieved information, leading to three critical challenges: determining when to retrieve, what modality to incorporate, and how to utilize retrieved information effectively. To address these challenges, we introduce Windsock, a query-dependent module making decisions on retrieval necessity and modality selection, effectively reducing computational overhead and improving response quality. Additionally, we propose Dynamic Noise-Resistance (DANCE) Instruction Tuning, an adaptive training strategy that enhances MLLMs' ability to utilize retrieved information while maintaining robustness against noise. Moreover, we adopt a self-assessment approach leveraging knowledge within MLLMs to convert question-answering datasets to MRAG training datasets. Extensive experiments demonstrate that our proposed method significantly improves the generation quality by 17.07% while reducing 8.95% retrieval times.

Yusen Zhang, Wenliang Zheng, Aashrith Madasu et al.

High-resolution image (HRI) understanding aims to process images with a large number of pixels, such as pathological images and agricultural aerial images, both of which can exceed 1 million pixels. Vision Large Language Models (VLMs) can allegedly handle HRIs, however, there is a lack of a comprehensive benchmark for VLMs to evaluate HRI understanding. To address this gap, we introduce HRScene, a novel unified benchmark for HRI understanding with rich scenes. HRScene incorporates 25 real-world datasets and 2 synthetic diagnostic datasets with resolutions ranging from 1,024 $\times$ 1,024 to 35,503 $\times$ 26,627. HRScene is collected and re-annotated by 10 graduate-level annotators, covering 25 scenarios, ranging from microscopic to radiology images, street views, long-range pictures, and telescope images. It includes HRIs of real-world objects, scanned documents, and composite multi-image. The two diagnostic evaluation datasets are synthesized by combining the target image with the gold answer and distracting images in different orders, assessing how well models utilize regions in HRI. We conduct extensive experiments involving 28 VLMs, including Gemini 2.0 Flash and GPT-4o. Experiments on HRScene show that current VLMs achieve an average accuracy of around 50% on real-world tasks, revealing significant gaps in HRI understanding. Results on synthetic datasets reveal that VLMs struggle to effectively utilize HRI regions, showing significant Regional Divergence and lost-in-middle, shedding light on future research.

Shicheng Tan, Huanjing Zhao, Shu Zhao et al.

Hyperbolic Neural Networks (HNNs), operating in hyperbolic space, have been widely applied in recent years, motivated by the existence of an optimal embedding in hyperbolic space that can preserve data hierarchical relationships (termed Hierarchical Representation Capability, HRC) more accurately than Euclidean space. However, there is no evidence to suggest that HNNs can achieve this theoretical optimal embedding, leading to much research being built on flawed motivations. In this paper, we propose a benchmark for evaluating HRC and conduct a comprehensive analysis of why HNNs are effective through large-scale experiments. Inspired by the analysis results, we propose several pre-training strategies to enhance HRC and improve the performance of downstream tasks, further validating the reliability of the analysis. Experiments show that HNNs cannot achieve the theoretical optimal embedding. The HRC is significantly affected by the optimization objectives and hierarchical structures, and enhancing HRC through pre-training strategies can significantly improve the performance of HNNs.

Shu Zhao, Huijuan Xu

Humans inherently recognize objects via selective visual perception, transform specific regions from the visual field into structured symbolic knowledge, and reason their relationships among regions based on the allocation of limited attention resources in line with humans' goals. While it is intuitive for humans, contemporary perception systems falter in extracting structural information due to the intricate cognitive abilities and commonsense knowledge required. To fill this gap, we present a new task called Local Scene Graph Generation. Distinct from the conventional scene graph generation task, which encompasses generating all objects and relationships in an image, our proposed task aims to abstract pertinent structural information with partial objects and their relationships for boosting downstream tasks that demand advanced comprehension and reasoning capabilities. Correspondingly, we introduce zEro-shot Local scEne GrAph geNeraTion (ELEGANT), a framework harnessing foundation models renowned for their powerful perception and commonsense reasoning, where collaboration and information communication among foundation models yield superior outcomes and realize zero-shot local scene graph generation without requiring labeled supervision. Furthermore, we propose a novel open-ended evaluation metric, Entity-level CLIPScorE (ECLIPSE), surpassing previous closed-set evaluation metrics by transcending their limited label space, offering a broader assessment. Experiment results show that our approach markedly outperforms baselines in the open-ended evaluation setting, and it also achieves a significant performance boost of up to 24.58% over prior methods in the close-set setting, demonstrating the effectiveness and powerful reasoning ability of our proposed framework.

Shicheng Tan, Shu Zhao, Yanping Zhang

Distributed document representation is one of the basic problems in natural language processing. Currently distributed document representation methods mainly consider the context information of words or sentences. These methods do not take into account the coherence of the document as a whole, e.g., a relation between the paper title and abstract, headline and description, or adjacent bodies in the document. The coherence shows whether a document is meaningful, both logically and syntactically, especially in scientific documents (papers or patents, etc.). In this paper, we propose a coupled text pair embedding (CTPE) model to learn the representation of scientific documents, which maintains the coherence of the document with coupled text pairs formed by segmenting the document. First, we divide the document into two parts (e.g., title and abstract, etc) which construct a coupled text pair. Then, we adopt negative sampling to construct uncoupled text pairs whose two parts are from different documents. Finally, we train the model to judge whether the text pair is coupled or uncoupled and use the obtained embedding of coupled text pairs as the embedding of documents. We perform experiments on three datasets for one information retrieval task and two recommendation tasks. The experimental results verify the effectiveness of the proposed CTPE model.

Jun Chen, Guang Yang, Habib Khan et al.

Automated and accurate segmentations of left atrium (LA) and atrial scars from late gadolinium-enhanced cardiac magnetic resonance (LGE CMR) images are in high demand for quantifying atrial scars. The previous quantification of atrial scars relies on a two-phase segmentation for LA and atrial scars due to their large volume difference (unbalanced atrial targets). In this paper, we propose an inter-cascade generative adversarial network, namely JAS-GAN, to segment the unbalanced atrial targets from LGE CMR images automatically and accurately in an end-to-end way. Firstly, JAS-GAN investigates an adaptive attention cascade to automatically correlate the segmentation tasks of the unbalanced atrial targets. The adaptive attention cascade mainly models the inclusion relationship of the two unbalanced atrial targets, where the estimated LA acts as the attention map to adaptively focus on the small atrial scars roughly. Then, an adversarial regularization is applied to the segmentation tasks of the unbalanced atrial targets for making a consistent optimization. It mainly forces the estimated joint distribution of LA and atrial scars to match the real ones. We evaluated the performance of our JAS-GAN on a 3D LGE CMR dataset with 192 scans. Compared with the state-of-the-art methods, our proposed approach yielded better segmentation performance (Average Dice Similarity Coefficient (DSC) values of 0.946 and 0.821 for LA and atrial scars, respectively), which indicated the effectiveness of our proposed approach for segmenting unbalanced atrial targets.

Shu Zhao, Dayan Wu, Yucan Zhou et al.

Deep hashing methods have shown great retrieval accuracy and efficiency in large-scale image retrieval. How to optimize discrete hash bits is always the focus in deep hashing methods. A common strategy in these methods is to adopt an activation function, e.g. $\operatorname{sigmoid}(\cdot)$ or $\operatorname{tanh}(\cdot)$, and minimize a quantization loss to approximate discrete values. However, this paradigm may make more and more hash bits stuck into the wrong saturated area of the activation functions and never escaped. We call this problem "Dead Bits Problem~(DBP)". Besides, the existing quantization loss will aggravate DBP as well. In this paper, we propose a simple but effective gradient amplifier which acts before activation functions to alleviate DBP. Moreover, we devise an error-aware quantization loss to further alleviate DBP. It avoids the negative effect of quantization loss based on the similarity between two images. The proposed gradient amplifier and error-aware quantization loss are compatible with a variety of deep hashing methods. Experimental results on three datasets demonstrate the efficiency of the proposed gradient amplifier and the error-aware quantization loss.

Dong Zhang, Shu Zhao, Zhen Duan et al.

Paper-reviewer recommendation task is of significant academic importance for conference chairs and journal editors. How to effectively and accurately recommend reviewers for the submitted papers is a meaningful and still tough task. In this paper, we propose a Multi-Label Classification method using a hierarchical and transparent Representation named Hiepar-MLC. Further, we propose a simple multi-label-based reviewer assignment MLBRA strategy to select the appropriate reviewers. It is interesting that we also explore the paper-reviewer recommendation in the coarse-grained granularity.

Jun Chen, Heye Zhang, Yanping Zhang et al.

Deep learning based task systems normally rely on a large amount of manually labeled training data, which is expensive to obtain and subject to operator variations. Moreover, it does not always hold that the manually labeled data and the unlabeled data are sitting in the same distribution. In this paper, we alleviate these problems by proposing a discriminative consistent domain generation (DCDG) approach to achieve a semi-supervised learning. The discriminative consistent domain is achieved by a double-sided domain adaptation. The double-sided domain adaptation aims to make a fusion of the feature spaces of labeled data and unlabeled data. In this way, we can fit the differences of various distributions between labeled data and unlabeled data. In order to keep the discriminativeness of generated consistent domain for the task learning, we apply an indirect learning for the double-sided domain adaptation. Based on the generated discriminative consistent domain, we can use the unlabeled data to learn the task model along with the labeled data via a consistent image generation. We demonstrate the performance of our proposed DCDG on the late gadolinium enhancement cardiac MRI (LGE-CMRI) images acquired from patients with atrial fibrillation in two clinical centers for the segmentation of the left atrium anatomy (LA) and proximal pulmonary veins (PVs). The experiments show that our semi-supervised approach achieves compelling segmentation results, which can prove the robustness of DCDG for the semi-supervised learning using the unlabeled data along with labeled data acquired from a single center or multicenter studies.

Dong Zhang, Guang Yang, Shu Zhao et al.

The quantification of the coronary artery stenosis is of significant clinical importance in coronary artery disease diagnosis and intervention treatment. It aims to quantify the morphological indices of the coronary artery lesions such as minimum lumen diameter, reference vessel diameter, lesion length, and these indices are the reference of the interventional stent placement. In this study, we propose a direct multiview quantitative coronary angiography (DMQCA) model as an automatic clinical tool to quantify the coronary artery stenosis from X-ray coronary angiography images. The proposed DMQCA model consists of a multiview module with two attention mechanisms, a key-frame module, and a regression module, to achieve direct accurate multiple-index estimation. The multi-view module comprehensively learns the Spatio-temporal features of coronary arteries through a three-dimensional convolution. The attention mechanisms of each view focus on the subtle feature of the lesion region and capture the important context information. The key-frame module learns the subtle features of the stenosis through successive dilated residual blocks. The regression module finally generates the indices estimation from multiple features.

Chenchu Xu, Lei Xu, Zhifan Gao et al.

Accurate detection of the myocardial infarction (MI) area is crucial for early diagnosis planning and follow-up management. In this study, we propose an end-to-end deep-learning algorithm framework (OF-RNN ) to accurately detect the MI area at the pixel level. Our OF-RNN consists of three different function layers: the heart localization layers, which can accurately and automatically crop the region-of-interest (ROI) sequences, including the left ventricle, using the whole cardiac magnetic resonance image sequences; the motion statistical layers, which are used to build a time-series architecture to capture two types of motion features (at the pixel-level) by integrating the local motion features generated by long short-term memory-recurrent neural networks and the global motion features generated by deep optical flows from the whole ROI sequence, which can effectively characterize myocardial physiologic function; and the fully connected discriminate layers, which use stacked auto-encoders to further learn these features, and they use a softmax classifier to build the correspondences from the motion features to the tissue identities (infarction or not) for each pixel. Through the seamless connection of each layer, our OF-RNN can obtain the area, position, and shape of the MI for each patient. Our proposed framework yielded an overall classification accuracy of 94.35% at the pixel level, from 114 clinical subjects. These results indicate the potential of our proposed method in aiding standardized MI assessments.