The rapid growth of big data and advancements in computational techniques have significantly transformed sports analytics. However, the diverse range of data sources -- including structured statistics, semi-structured formats like sensor data, and unstructured media such as written articles, audio, and video -- creates substantial challenges in extracting actionable insights. These various formats, often referred to as multimodal data, require integration to fully leverage their potential. Conventional systems, which typically prioritize structured data, face limitations when processing and combining these diverse content types, reducing their effectiveness in real-time sports analysis. To address these challenges, recent research highlights the importance of multimodal data integration for capturing the complexity of real-world sports environments. Building on this foundation, this paper introduces GridMind, a multi-agent framework that unifies structured, semi-structured, and unstructured data through Retrieval-Augmented Generation (RAG) and large language models (LLMs) to facilitate natural language querying of NFL data. This approach aligns with the evolving field of multimodal representation learning, where unified models are increasingly essential for real-time, cross-modal interactions. GridMind's distributed architecture includes specialized agents that autonomously manage each stage of a prompt -- from interpretation and data retrieval to response synthesis. This modular design enables flexible, scalable handling of multimodal data, allowing users to pose complex, context-rich questions and receive comprehensive, intuitive responses via a conversational interface.
Graph Transformers (GTs), which simultaneously integrate message-passing and self-attention mechanisms, have achieved promising empirical results in some graph prediction tasks. Although these approaches show the potential of Transformers in capturing long-range graph topology information, issues concerning the quadratic complexity and high computing energy consumption severely limit the scalability of GTs on large-scale graphs. Recently, as brain-inspired neural networks, Spiking Neural Networks (SNNs), facilitate the development of graph representation learning methods with lower computational and storage overhead through the unique event-driven spiking neurons. Inspired by these characteristics, we propose a linear-time Graph Transformer using Spiking Vector Quantization (GT-SVQ) for node classification. GT-SVQ reconstructs codebooks based on rate coding outputs from spiking neurons, and injects the codebooks into self-attention blocks to aggregate global information in linear complexity. Besides, spiking vector quantization effectively alleviates codebook collapse and the reliance on complex machinery (distance measure, auxiliary loss, etc.) present in previous vector quantization-based graph learning methods. In experiments, we compare GT-SVQ with other state-of-the-art baselines on node classification datasets ranging from small to large. Experimental results show that GT-SVQ has achieved competitive performances on most datasets while maintaining up to 130x faster inference speed compared to other GTs.
RGB-Thermal Video Object Detection (RGBT VOD) can address the limitation of traditional RGB-based VOD in challenging lighting conditions, making it more practical and effective in many applications. However, similar to most RGBT fusion tasks, it still mainly relies on manually aligned multimodal image pairs. In this paper, we propose a novel Multimodal Spatio-temporal Graph learning Network (MSGNet) for alignment-free RGBT VOD problem by leveraging the robust graph representation learning model. Specifically, we first design an Adaptive Partitioning Layer (APL) to estimate the corresponding regions of the Thermal image within the RGB image (high-resolution), achieving a preliminary inexact alignment. Then, we introduce the Spatial Sparse Graph Learning Module (S-SGLM) which employs a sparse information passing mechanism on the estimated inexact alignment to achieve reliable information interaction between different modalities. Moreover, to fully exploit the temporal cues for RGBT VOD problem, we introduce Hybrid Structured Temporal Modeling (HSTM), which involves a Temporal Sparse Graph Learning Module (T-SGLM) and Temporal Star Block (TSB). T-SGLM aims to filter out some redundant information between adjacent frames by employing the sparse aggregation mechanism on the temporal graph. Meanwhile, TSB is dedicated to achieving the complementary learning of local spatial relationships. Extensive comparative experiments conducted on both the aligned dataset VT-VOD50 and the unaligned dataset UVT-VOD2024 demonstrate the effectiveness and superiority of our proposed method. Our project will be made available on our website for free public access.
Vision foundation models in remote sensing have been extensively studied due to their superior generalization on various downstream tasks. Synthetic Aperture Radar (SAR) offers all-day, all-weather imaging capabilities, providing significant advantages for Earth observation. However, establishing a foundation model for SAR image interpretation inevitably encounters the challenges of insufficient information utilization and poor interpretability. In this paper, we propose a remote sensing foundation model based on complex-valued SAR data, which simulates the polarimetric decomposition process for pre-training, i.e., characterizing pixel scattering intensity as a weighted combination of scattering bases and scattering coefficients, thereby endowing the foundation model with physical interpretability. Specifically, we construct a series of scattering queries, each representing an independent and meaningful scattering basis, which interact with SAR features in the scattering query decoder and output the corresponding scattering coefficient. To guide the pre-training process, polarimetric decomposition loss and power self-supervision loss are constructed. The former aligns the predicted coefficients with Yamaguchi coefficients, while the latter reconstructs power from the predicted coefficients and compares it to the input image's power. The performance of our foundation model is validated on six typical downstream tasks, achieving state-of-the-art results. Notably, the foundation model can extract stable feature representations and exhibits strong generalization, even in data-scarce conditions.
Spatiotemporal learning is challenging due to the intricate interplay between spatial and temporal dependencies, the high dimensionality of the data, and scalability constraints. These challenges are further amplified in scientific domains, where data is often irregularly distributed (e.g., missing values from sensor failures) and high-volume (e.g., high-fidelity simulations), posing additional computational and modeling difficulties. In this paper, we present SCENT, a novel framework for scalable and continuity-informed spatiotemporal representation learning. SCENT unifies interpolation, reconstruction, and forecasting within a single architecture. Built on a transformer-based encoder-processor-decoder backbone, SCENT introduces learnable queries to enhance generalization and a query-wise cross-attention mechanism to effectively capture multi-scale dependencies. To ensure scalability in both data size and model complexity, we incorporate a sparse attention mechanism, enabling flexible output representations and efficient evaluation at arbitrary resolutions. We validate SCENT through extensive simulations and real-world experiments, demonstrating state-of-the-art performance across multiple challenging tasks while achieving superior scalability.
Humans can develop internal world models that encode common sense knowledge, telling them how the world works and predicting the consequences of their actions. This concept has emerged as a promising direction for establishing general-purpose machine-learning models in recent preliminary works, e.g., for visual representation learning. In this paper, we present CheXWorld, the first effort towards a self-supervised world model for radiographic images. Specifically, our work develops a unified framework that simultaneously models three aspects of medical knowledge essential for qualified radiologists, including 1) local anatomical structures describing the fine-grained characteristics of local tissues (e.g., architectures, shapes, and textures); 2) global anatomical layouts describing the global organization of the human body (e.g., layouts of organs and skeletons); and 3) domain variations that encourage CheXWorld to model the transitions across different appearance domains of radiographs (e.g., varying clarity, contrast, and exposure caused by collecting radiographs from different hospitals, devices, or patients). Empirically, we design tailored qualitative and quantitative analyses, revealing that CheXWorld successfully captures these three dimensions of medical knowledge. Furthermore, transfer learning experiments across eight medical image classification and segmentation benchmarks showcase that CheXWorld significantly outperforms existing SSL methods and large-scale medical foundation models. Code & pre-trained models are available at https://github.com/LeapLabTHU/CheXWorld.
Recent works in multimodal recommendations, which leverage diverse modal information to address data sparsity and enhance recommendation accuracy, have garnered considerable interest. Two key processes in multimodal recommendations are modality fusion and representation learning. Previous approaches in modality fusion often employ simplistic attentive or pre-defined strategies at early or late stages, failing to effectively handle irrelevant information among modalities. In representation learning, prior research has constructed heterogeneous and homogeneous graph structures encapsulating user-item, user-user, and item-item relationships to better capture user interests and item profiles. Modality fusion and representation learning were considered as two independent processes in previous work. In this paper, we reveal that these two processes are complementary and can support each other. Specifically, powerful representation learning enhances modality fusion, while effective fusion improves representation quality. Stemming from these two processes, we introduce a COmposite grapH convolutional nEtwork with dual-stage fuSION for the multimodal recommendation, named COHESION. Specifically, it introduces a dual-stage fusion strategy to reduce the impact of irrelevant information, refining all modalities using ID embedding in the early stage and fusing their representations at the late stage. It also proposes a composite graph convolutional network that utilizes user-item, user-user, and item-item graphs to extract heterogeneous and homogeneous latent relationships within users and items. Besides, it introduces a novel adaptive optimization to ensure balanced and reasonable representations across modalities. Extensive experiments on three widely used datasets demonstrate the significant superiority of COHESION over various competitive baselines.
Efficiently transferring Learned Image Compression (LIC) model from human perception to machine perception is an emerging challenge in vision-centric representation learning. Existing approaches typically adapt LIC to downstream tasks in a single-task manner, which is inefficient, lacks task interaction, and results in multiple task-specific bitstreams. To address these limitations, we propose an asymmetric adaptor framework that supports multi-task adaptation within a single model. Our method introduces a shared adaptor to learn general semantic features and task-specific adaptors to preserve task-level distinctions. With only lightweight plug-in modules and a frozen base codec, our method achieves strong performance across multiple tasks while maintaining compression efficiency. Experiments on the PASCAL-Context benchmark demonstrate that our method outperforms both Fully Fine-Tuned and other Parameter Efficient Fine-Tuned (PEFT) baselines, and validating the effectiveness of multi-vision transferring.
After pre-training on extensive image-text pairs, Contrastive Language-Image Pre-training (CLIP) demonstrates promising performance on a wide variety of benchmarks. However, a substantial volume of multimodal interleaved documents remains underutilized for contrastive vision-language representation learning. To fully leverage these unpaired documents, we initially establish a Real-World Data Extraction pipeline to extract high-quality images and texts. Then we design a hierarchical retrieval method to efficiently associate each image with multiple semantically relevant realistic texts. To further enhance fine-grained visual information, we propose an image semantic augmented generation module for synthetic text production. Furthermore, we employ a semantic balance sampling strategy to improve dataset diversity, enabling better learning of long-tail concepts. Based on these innovations, we construct RealSyn, a dataset combining realistic and synthetic texts, available in three scales: 15M, 30M, and 100M. We compare our dataset with other widely used datasets of equivalent scale for CLIP training. Models pre-trained on RealSyn consistently achieve state-of-the-art performance across various downstream tasks, including linear probe, zero-shot transfer, zero-shot robustness, and zero-shot retrieval. Furthermore, extensive experiments confirm that RealSyn significantly enhances contrastive vision-language representation learning and demonstrates robust scalability. To facilitate future research, the RealSyn dataset and pretrained model weights are released at https://github.com/deepglint/RealSyn.
Contrastive video-language pretraining has demonstrated great success in learning rich and robust video representations. However, deploying such video encoders on compute-constrained edge devices remains challenging due to their high computational demands. Additionally, existing models are typically trained to process only short video clips, often limited to 4 to 64 frames. In this paper, we introduce AdaVid, a flexible architectural framework designed to learn efficient video encoders that can dynamically adapt their computational footprint based on available resources. At the heart of AdaVid is an adaptive transformer block, inspired by Matryoshka Representation Learning, which allows the model to adjust its hidden embedding dimension at inference time. We show that AdaVid-EgoVLP, trained on video-narration pairs from the large-scale Ego4D dataset, matches the performance of the standard EgoVLP on short video-language benchmarks using only half the compute, and even outperforms EgoVLP when given equal computational resources. We further explore the trade-off between frame count and compute on the challenging Diving48 classification benchmark, showing that AdaVid enables the use of more frames without exceeding computational limits. To handle longer videos, we also propose a lightweight hierarchical network that aggregates short clip features, achieving a strong balance between compute efficiency and accuracy across several long video benchmarks.
Evidence-based deep learning represents a burgeoning paradigm for uncertainty estimation, offering reliable predictions with negligible extra computational overheads. Existing methods usually adopt Kullback-Leibler divergence to estimate the uncertainty of network predictions, ignoring domain gaps among various modalities. To tackle this issue, this paper introduces a novel algorithm based on H\"older Divergence (HD) to enhance the reliability of multi-view learning by addressing inherent uncertainty challenges from incomplete or noisy data. Generally, our method extracts the representations of multiple modalities through parallel network branches, and then employs HD to estimate the prediction uncertainties. Through the Dempster-Shafer theory, integration of uncertainty from different modalities, thereby generating a comprehensive result that considers all available representations. Mathematically, HD proves to better measure the ``distance'' between real data distribution and predictive distribution of the model and improve the performances of multi-class recognition tasks. Specifically, our method surpass the existing state-of-the-art counterparts on all evaluating benchmarks. We further conduct extensive experiments on different backbones to verify our superior robustness. It is demonstrated that our method successfully pushes the corresponding performance boundaries. Finally, we perform experiments on more challenging scenarios, \textit{i.e.}, learning with incomplete or noisy data, revealing that our method exhibits a high tolerance to such corrupted data.
Text-attributed graphs (TAGs) present unique challenges in representation learning by requiring models to capture both the semantic richness of node-associated texts and the structural dependencies of the graph. While graph neural networks (GNNs) excel at modeling topological information, they lack the capacity to process unstructured text. Conversely, large language models (LLMs) are proficient in text understanding but are typically unaware of graph structure. In this work, we propose BiGTex (Bidirectional Graph Text), a novel architecture that tightly integrates GNNs and LLMs through stacked Graph-Text Fusion Units. Each unit allows for mutual attention between textual and structural representations, enabling information to flow in both directions, text influencing structure and structure guiding textual interpretation. The proposed architecture is trained using parameter-efficient fine-tuning (LoRA), keeping the LLM frozen while adapting to task-specific signals. Extensive experiments on five benchmark datasets demonstrate that BiGTex achieves state-of-the-art performance in node classification and generalizes effectively to link prediction. An ablation study further highlights the importance of soft prompting and bi-directional attention in the model's success.
Recent developments in generative artificial intelligence (AI) rely on machine learning techniques such as deep learning and generative modeling to achieve state-of-the-art performance across wide-ranging domains. These methods' surprising performance is due in part to their ability to learn implicit "representations'' of complex, multi-modal data. Unfortunately, deep neural networks are notoriously black boxes that obscure these representations, making them difficult to interpret or analyze. To resolve these difficulties, one approach is to build new interpretable neural network models from the ground up. This is the goal of the emerging field of causal representation learning (CRL) that uses causality as a vector for building flexible, interpretable, and transferable generative AI. CRL can be seen as a culmination of three intrinsically statistical problems: (i) latent variable models such as factor analysis; (ii) causal graphical models with latent variables; and (iii) nonparametric statistics and deep learning. This paper reviews recent progress in CRL from a statistical perspective, focusing on connections to classical models and statistical and causal identifiablity results. This review also highlights key application areas, implementation strategies, and open statistical questions in CRL.
Graph Neural Networks (GNNs) struggle to balance heterophily and homophily in representation learning, a challenge further amplified in self-supervised settings. We propose H$^3$GNNs, an end-to-end self-supervised learning framework that harmonizes both structural properties through two key innovations: (i) Joint Structural Node Encoding. We embed nodes into a unified space combining linear and non-linear feature projections with K-hop structural representations via a Weighted Graph Convolution Network(WGCN). A cross-attention mechanism enhances awareness and adaptability to heterophily and homophily. (ii) Self-Supervised Learning Using Teacher-Student Predictive Architectures with Node-Difficulty Driven Dynamic Masking Strategies. We use a teacher-student model, the student sees the masked input graph and predicts node features inferred by the teacher that sees the full input graph in the joint encoding space. To enhance learning difficulty, we introduce two novel node-predictive-difficulty-based masking strategies. Experiments on seven benchmarks (four heterophily datasets and three homophily datasets) confirm the effectiveness and efficiency of H$^3$GNNs across diverse graph types. Our H$^3$GNNs achieves overall state-of-the-art performance on the four heterophily datasets, while retaining on-par performance to previous state-of-the-art methods on the three homophily datasets.
Multimodal foundation models have significantly improved feature representation by integrating information from multiple modalities, making them highly suitable for a broader set of applications. However, the exploration of multimodal facial representation for understanding perception has been limited. Understanding and analyzing facial states, such as Action Units (AUs) and emotions, require a comprehensive and robust framework that bridges visual and linguistic modalities. In this paper, we present a comprehensive pipeline for multimodal facial state analysis. First, we compile a new Multimodal Face Dataset (MFA) by generating detailed multilevel language descriptions of face, incorporating Action Unit (AU) and emotion descriptions, by leveraging GPT-4o. Second, we introduce a novel Multilevel Multimodal Face Foundation model (MF^2) tailored for Action Unit (AU) and emotion recognition. Our model incorporates comprehensive visual feature modeling at both local and global levels of face image, enhancing its ability to represent detailed facial appearances. This design aligns visual representations with structured AU and emotion descriptions, ensuring effective cross-modal integration. Third, we develop a Decoupled Fine-Tuning Network (DFN) that efficiently adapts MF^2 across various tasks and datasets. This approach not only reduces computational overhead but also broadens the applicability of the foundation model to diverse scenarios. Experimentation show superior performance for AU and emotion detection tasks.
The availability of large, structured populations of neural networks - called 'model zoos' - has led to the development of a multitude of downstream tasks ranging from model analysis, to representation learning on model weights or generative modeling of neural network parameters. However, existing model zoos are limited in size and architecture and neglect the transformer, which is among the currently most successful neural network architectures. We address this gap by introducing the first model zoo of vision transformers (ViT). To better represent recent training approaches, we develop a new blueprint for model zoo generation that encompasses both pre-training and fine-tuning steps, and publish 250 unique models. They are carefully generated with a large span of generating factors, and their diversity is validated using a thorough choice of weight-space and behavioral metrics. To further motivate the utility of our proposed dataset, we suggest multiple possible applications grounded in both extensive exploratory experiments and a number of examples from the existing literature. By extending previous lines of similar work, our model zoo allows researchers to push their model population-based methods from the small model regime to state-of-the-art architectures. We make our model zoo available at github.com/ModelZoos/ViTModelZoo.
Multimodal representation learning, exemplified by multimodal contrastive learning (MMCL) using image-text pairs, aims to learn powerful representations by aligning cues across modalities. This approach relies on the core assumption that the exemplar image-text pairs constitute two representations of an identical concept. However, recent research has revealed that real-world datasets often exhibit cross-modal misalignment. There are two distinct viewpoints on how to address this issue: one suggests mitigating the misalignment, and the other leveraging it. We seek here to reconcile these seemingly opposing perspectives, and to provide a practical guide for practitioners. Using latent variable models we thus formalize cross-modal misalignment by introducing two specific mechanisms: Selection bias, where some semantic variables are absent in the text, and perturbation bias, where semantic variables are altered -- both leading to misalignment in data pairs. Our theoretical analysis demonstrates that, under mild assumptions, the representations learned by MMCL capture exactly the information related to the subset of the semantic variables invariant to selection and perturbation biases. This provides a unified perspective for understanding misalignment. Based on this, we further offer actionable insights into how misalignment should inform the design of real-world ML systems. We validate our theoretical findings via extensive empirical studies on both synthetic data and real image-text datasets, shedding light on the nuanced impact of cross-modal misalignment on multimodal representation learning.
Out-of-distribution (OOD) detection is essential for the safe deployment of machine learning models. Recent advances have explored improved classification losses and representation learning strategies to enhance OOD detection. However, these methods are often tailored to specific post-hoc detection techniques, limiting their generalizability. In this work, we identify a critical issue in Logit Normalization (LogitNorm), which inhibits its effectiveness in improving certain post-hoc OOD detection methods. To address this, we propose Extended Logit Normalization ($\textbf{ELogitNorm}$), a novel hyperparameter-free formulation that significantly benefits a wide range of post-hoc detection methods. By incorporating feature distance-awareness to LogitNorm, $\textbf{ELogitNorm}$ shows more robust OOD separability and in-distribution (ID) confidence calibration than its predecessor. Extensive experiments across standard benchmarks demonstrate that our approach outperforms state-of-the-art training-time methods in OOD detection while maintaining strong ID classification accuracy.
The rapid accumulation of Electronic Health Records (EHRs) has transformed healthcare by providing valuable data that enhance clinical predictions and diagnoses. While conventional machine learning models have proven effective, they often lack robust representation learning and depend heavily on expert-crafted features. Although deep learning offers powerful solutions, it is often criticized for its lack of interpretability. To address these challenges, we propose DeepSelective, a novel end to end deep learning framework for predicting patient prognosis using EHR data, with a strong emphasis on enhancing model interpretability. DeepSelective combines data compression techniques with an innovative feature selection approach, integrating custom-designed modules that work together to improve both accuracy and interpretability. Our experiments demonstrate that DeepSelective not only enhances predictive accuracy but also significantly improves interpretability, making it a valuable tool for clinical decision-making. The source code is freely available at http://www.healthinformaticslab.org/supp/resources.php .
Recommending items to users has long been a fundamental task, and studies have tried to improve it ever since. Most well-known models commonly employ representation learning to map users and items into a unified embedding space for matching assessment. These approaches have primary limitations, especially when dealing with explicit feedback and sparse data contexts. Two primary limitations are their proneness to overfitting and failure to incorporate epistemic uncertainty in predictions. To address these problems, we propose a novel Bayesian Deep Ensemble Collaborative Filtering method named BDECF. To improve model generalization and quality, we utilize Bayesian Neural Networks, which incorporate uncertainty within their weight parameters. In addition, we introduce a new interpretable non-linear matching approach for the user and item embeddings, leveraging the advantages of the attention mechanism. Furthermore, we endorse the implementation of an ensemble-based supermodel to generate more robust and reliable predictions, resulting in a more complete model. Empirical evaluation through extensive experiments and ablation studies across a range of publicly accessible real-world datasets with differing sparsity characteristics confirms our proposed method's effectiveness and the importance of its components.