Invertible neural architectures have recently attracted attention for their compactness, interpretability, and information-preserving properties. In this work, we propose the Fourier-Invertible Neural Encoder (FINE), which combines invertible monotonic activation functions with reversible filter structures, and could be extended using Invertible ResNets. This architecture is examined in learning low-dimensional representations of one-dimensional nonlinear wave interactions and exact circular translation symmetry. Dimensionality is preserved across layers, except for a Fourier truncation step in the latent space, which enables dimensionality reduction while maintaining shift equivariance and interpretability. Our results demonstrate that FINE significantly outperforms classical linear methods such as Discrete Fourier Transformation (DFT) and Proper Orthogonal Decomposition (POD), and achieves reconstruction accuracy better than conventional deep autoencoders with convolutional layers (CNN) - while using substantially smaller models and offering superior physical interpretability. These findings suggest that invertible single-neuron networks, when combined with spectral truncation, offer a promising framework for learning compact and interpretable representations of physics datasets, and symmetry-aware representation learning in physics-informed machine learning.
Omni-domain infrared small target detection (IRSTD) poses formidable challenges, as a single model must seamlessly adapt to diverse imaging systems, varying resolutions, and multiple spectral bands simultaneously. Current approaches predominantly rely on visual-only modeling paradigms that not only struggle with complex background interference and inherently scarce target features, but also exhibit limited generalization capabilities across complex omni-scene environments where significant domain shifts and appearance variations occur. In this work, we reveal a critical oversight in existing paradigms: the neglect of readily available auxiliary metadata describing imaging parameters and acquisition conditions, such as spectral bands, sensor platforms, resolution, and observation perspectives. To address this limitation, we propose the Auxiliary Metadata Driven Infrared Small Target Detector (AuxDet), a novel multi-modal framework that fundamentally reimagines the IRSTD paradigm by incorporating textual metadata for scene-aware optimization. Through a high-dimensional fusion module based on multi-layer perceptrons (MLPs), AuxDet dynamically integrates metadata semantics with visual features, guiding adaptive representation learning for each individual sample. Additionally, we design a lightweight prior-initialized enhancement module using 1D convolutional blocks to further refine fused features and recover fine-grained target cues. Extensive experiments on the challenging WideIRSTD-Full benchmark demonstrate that AuxDet consistently outperforms state-of-the-art methods, validating the critical role of auxiliary information in improving robustness and accuracy in omni-domain IRSTD tasks. Code is available at https://github.com/GrokCV/AuxDet.
Generalized gait recognition, which aims to achieve robust performance across diverse domains, remains a challenging problem due to severe domain shifts in viewpoints, appearances, and environments. While mixed-dataset training is widely used to enhance generalization, it introduces new obstacles including inter-dataset optimization conflicts and redundant or noisy samples, both of which hinder effective representation learning. To address these challenges, we propose a unified framework that systematically improves cross-domain gait recognition. First, we design a disentangled triplet loss that isolates supervision signals across datasets, mitigating gradient conflicts during optimization. Second, we introduce a targeted dataset distillation strategy that filters out the least informative 20\% of training samples based on feature redundancy and prediction uncertainty, enhancing data efficiency. Extensive experiments on CASIA-B, OU-MVLP, Gait3D, and GREW demonstrate that our method significantly improves cross-dataset recognition for both GaitBase and DeepGaitV2 backbones, without sacrificing source-domain accuracy. Code will be released at https://github.com/li1er3/Generalized_Gait.
Semi-supervised learning (SSL) has achieved significant progress in medical image segmentation (SSMIS) through effective utilization of limited labeled data. While current SSL methods for medical images predominantly rely on consistency regularization and pseudo-labeling, they often overlook transferable semantic relationships across different clinical domains and imaging modalities. To address this, we propose TransMedSeg, a novel transferable semantic framework for semi-supervised medical image segmentation. Our approach introduces a Transferable Semantic Augmentation (TSA) module, which implicitly enhances feature representations by aligning domain-invariant semantics through cross-domain distribution matching and intra-domain structural preservation. Specifically, TransMedSeg constructs a unified feature space where teacher network features are adaptively augmented towards student network semantics via a lightweight memory module, enabling implicit semantic transformation without explicit data generation. Interestingly, this augmentation is implicitly realized through an expected transferable cross-entropy loss computed over the augmented teacher distribution. An upper bound of the expected loss is theoretically derived and minimized during training, incurring negligible computational overhead. Extensive experiments on medical image datasets demonstrate that TransMedSeg outperforms existing semi-supervised methods, establishing a new direction for transferable representation learning in medical image analysis.
Motion forecasting represents a critical challenge in autonomous driving systems, requiring accurate prediction of surrounding agents' future trajectories. While existing approaches predict future motion states with the extracted scene context feature from historical agent trajectories and road layouts, they suffer from the information degradation during the scene feature encoding. To address the limitation, we propose HAMF, a novel motion forecasting framework that learns future motion representations with the scene context encoding jointly, to coherently combine the scene understanding and future motion state prediction. We first embed the observed agent states and map information into 1D token sequences, together with the target multi-modal future motion features as a set of learnable tokens. Then we design a unified Attention-based encoder, which synergistically combines self-attention and cross-attention mechanisms to model the scene context information and aggregate future motion features jointly. Complementing the encoder, we implement the Mamba module in the decoding stage to further preserve the consistency and correlations among the learned future motion representations, to generate the accurate and diverse final trajectories. Extensive experiments on Argoverse 2 benchmark demonstrate that our hybrid Attention-Mamba model achieves state-of-the-art motion forecasting performance with the simple and lightweight architecture.
With recent breakthroughs in large-scale modeling, the Segment Anything Model (SAM) has demonstrated significant potential in a variety of visual applications. However, due to the lack of underwater domain expertise, SAM and its variants face performance limitations in end-to-end underwater instance segmentation tasks, while their higher computational requirements further hinder their application in underwater scenarios. To address this challenge, we propose a large-scale underwater instance segmentation dataset, UIIS10K, which includes 10,048 images with pixel-level annotations for 10 categories. Then, we introduce UWSAM, an efficient model designed for automatic and accurate segmentation of underwater instances. UWSAM efficiently distills knowledge from the SAM ViT-Huge image encoder into the smaller ViT-Small image encoder via the Mask GAT-based Underwater Knowledge Distillation (MG-UKD) method for effective visual representation learning. Furthermore, we design an End-to-end Underwater Prompt Generator (EUPG) for UWSAM, which automatically generates underwater prompts instead of explicitly providing foreground points or boxes as prompts, thus enabling the network to locate underwater instances accurately for efficient segmentation. Comprehensive experimental results show that our model is effective, achieving significant performance improvements over state-of-the-art methods on multiple underwater instance datasets. Datasets and codes are available at https://github.com/LiamLian0727/UIIS10K.
Traditional time series models are task-specific and often depend on dataset-specific training and extensive feature engineering. While Transformer-based architectures have improved scalability, foundation models, commonplace in text, vision, and audio, remain under-explored for time series and are largely restricted to forecasting. We introduce $\textbf{CHARM}$, a foundation embedding model for multivariate time series that learns shared, transferable, and domain-aware representations. To address the unique difficulties of time series foundation learning, $\textbf{CHARM}$ incorporates architectural innovations that integrate channel-level textual descriptions while remaining invariant to channel order. The model is trained using a Joint Embedding Predictive Architecture (JEPA), with novel augmentation schemes and a loss function designed to improve interpretability and training stability. Our $7$M-parameter model achieves state-of-the-art performance across diverse downstream tasks, setting a new benchmark for time series representation learning.
The open-world assumption in model development suggests that a model might lack sufficient information to adequately handle data that is entirely distinct or out of distribution (OOD). While deep learning methods have shown promising results in handling OOD data through generalization techniques, they often require specialized hardware that may not be accessible to all users. We present TCL, a lightweight yet effective solution that operates efficiently on standard CPU hardware. Our approach adapts contrastive learning principles specifically for tabular data structures, incorporating full matrix augmentation and simplified loss calculation. Through comprehensive experiments across 10 diverse datasets, we demonstrate that TCL outperforms existing models, including FT-Transformer and ResNet, particularly in classification tasks, while maintaining competitive performance in regression problems. TCL achieves these results with significantly reduced computational requirements, making it accessible to users with limited hardware capabilities. This study also provides practical guidance for detecting and evaluating OOD data through straightforward experiments and visualizations. Our findings show that TCL offers a promising balance between performance and efficiency in handling OOD prediction tasks, which is particularly beneficial for general machine learning practitioners working with computational constraints.
With the rapid advancement of unmanned aerial vehicles (UAVs) and missile technologies, perimeter-defense game between attackers and defenders for the protection of critical regions have become increasingly complex and strategically significant across a wide range of domains. However, existing studies predominantly focus on small-scale, simplified two-dimensional scenarios, often overlooking realistic environmental perturbations, motion dynamics, and inherent heterogeneity--factors that pose substantial challenges to real-world applicability. To bridge this gap, we investigate large-scale heterogeneous perimeter-defense game in a three-dimensional setting, incorporating realistic elements such as motion dynamics and wind fields. We derive the Nash equilibrium strategies for both attackers and defenders, characterize the victory regions, and validate our theoretical findings through extensive simulations. To tackle large-scale heterogeneous control challenges in defense strategies, we propose an Embedded Mean-Field Actor-Critic (EMFAC) framework. EMFAC leverages representation learning to enable high-level action aggregation in a mean-field manner, supporting scalable coordination among defenders. Furthermore, we introduce a lightweight agent-level attention mechanism based on reward representation, which selectively filters observations and mean-field information to enhance decision-making efficiency and accelerate convergence in large-scale tasks. Extensive simulations across varying scales demonstrate the effectiveness and adaptability of EMFAC, which outperforms established baselines in both convergence speed and overall performance. To further validate practicality, we test EMFAC in small-scale real-world experiments and conduct detailed analyses, offering deeper insights into the framework's effectiveness in complex scenarios.
Despite widespread adoption, multimodal large language models (MLLMs) suffer performance degradation when encountering unfamiliar queries under distribution shifts. Existing methods to improve MLLM generalization typically require either more instruction data or larger advanced model architectures, both of which incur non-trivial human labor or computational costs. In this work, we take an alternative approach to enhance the robustness of MLLMs under distribution shifts, from a representation learning perspective. Inspired by the information bottleneck (IB) principle, we derive a variational lower bound of the IB for MLLMs and devise a practical implementation, Visual Instruction Bottleneck Tuning (Vittle). We then provide a theoretical justification of Vittle by revealing its connection to an information-theoretic robustness metric of MLLM. Empirical validation of three MLLMs on open-ended and closed-form question answering and object hallucination detection tasks over 45 datasets, including 30 shift scenarios, demonstrates that Vittle consistently improves the MLLM's robustness under shifts by pursuing the learning of a minimal sufficient representation.
Self-supervised representation learning for point cloud videos remains a challenging problem with two key limitations: (1) existing methods rely on explicit knowledge to learn motion, resulting in suboptimal representations; (2) prior Masked AutoEncoder (MAE) frameworks struggle to bridge the gap between low-level geometry and high-level dynamics in 4D data. In this work, we propose a novel self-disentangled MAE for learning expressive, discriminative, and transferable 4D representations. To overcome the first limitation, we learn motion by aligning high-level semantics in the latent space \textit{without any explicit knowledge}. To tackle the second, we introduce a \textit{self-disentangled learning} strategy that incorporates the latent token with the geometry token within a shared decoder, effectively disentangling low-level geometry and high-level semantics. In addition to the reconstruction objective, we employ three alignment objectives to enhance temporal understanding, including frame-level motion and video-level global information. We show that our pre-trained encoder surprisingly discriminates spatio-temporal representation without further fine-tuning. Extensive experiments on MSR-Action3D, NTU-RGBD, HOI4D, NvGesture, and SHREC'17 demonstrate the superiority of our approach in both coarse-grained and fine-grained 4D downstream tasks. Notably, Uni4D improves action segmentation accuracy on HOI4D by $+3.8\%$.
We introduce Perceptual-Initialization (PI), a paradigm shift in visual representation learning that incorporates human perceptual structure during the initialization phase rather than as a downstream fine-tuning step. By integrating human-derived triplet embeddings from the NIGHTS dataset to initialize a CLIP vision encoder, followed by self-supervised learning on YFCC15M, our approach demonstrates significant zero-shot performance improvements, without any task-specific fine-tuning, across 29 zero shot classification and 2 retrieval benchmarks. On ImageNet-1K, zero-shot gains emerge after approximately 15 epochs of pretraining. Benefits are observed across datasets of various scales, with improvements manifesting at different stages of the pretraining process depending on dataset characteristics. Our approach consistently enhances zero-shot top-1 accuracy, top-5 accuracy, and retrieval recall (e.g., R@1, R@5) across these diverse evaluation tasks, without requiring any adaptation to target domains. These findings challenge the conventional wisdom of using human-perceptual data primarily for fine-tuning and demonstrate that embedding human perceptual structure during early representation learning yields more capable and vision-language aligned systems that generalize immediately to unseen tasks. Our work shows that "beginning with you", starting with human perception, provides a stronger foundation for general-purpose vision-language intelligence.
Surgical phase recognition from video is a technology that automatically classifies the progress of a surgical procedure and has a wide range of potential applications, including real-time surgical support, optimization of medical resources, training and skill assessment, and safety improvement. Recent advances in surgical phase recognition technology have focused primarily on Transform-based methods, although methods that extract spatial features from individual frames using a CNN and video features from the resulting time series of spatial features using time series modeling have shown high performance. However, there remains a paucity of research on training methods for CNNs employed for feature extraction or representation learning in surgical phase recognition. In this study, we propose a method for representation learning in surgical workflow analysis using a vision-language model (ReSW-VL). Our proposed method involves fine-tuning the image encoder of a CLIP (Convolutional Language Image Model) vision-language model using prompt learning for surgical phase recognition. The experimental results on three surgical phase recognition datasets demonstrate the effectiveness of the proposed method in comparison to conventional methods.
Tactile perception is profoundly influenced by the surface properties of objects in contact. However, despite their crucial role in shaping tactile experiences, these material characteristics have been largely neglected in existing tactile representation learning methods. Most approaches primarily focus on aligning tactile data with visual or textual information, overlooking the richness of tactile feedback that comes from understanding the materials' inherent properties. In this work, we address this gap by revisiting the tactile representation learning framework and incorporating material-aware priors into the learning process. These priors, which represent pre-learned characteristics specific to different materials, allow tactile models to better capture and generalize the nuances of surface texture. Our method enables more accurate, contextually rich tactile feedback across diverse materials and textures, improving performance in real-world applications such as robotics, haptic feedback systems, and material editing.
We present Sat2Sound, a multimodal representation learning framework for soundscape mapping, designed to predict the distribution of sounds at any location on Earth. Existing methods for this task rely on satellite image and paired geotagged audio samples, which often fail to capture the diversity of sound sources at a given location. To address this limitation, we enhance existing datasets by leveraging a Vision-Language Model (VLM) to generate semantically rich soundscape descriptions for locations depicted in satellite images. Our approach incorporates contrastive learning across audio, audio captions, satellite images, and satellite image captions. We hypothesize that there is a fixed set of soundscape concepts shared across modalities. To this end, we learn a shared codebook of soundscape concepts and represent each sample as a weighted average of these concepts. Sat2Sound achieves state-of-the-art performance in cross-modal retrieval between satellite image and audio on two datasets: GeoSound and SoundingEarth. Additionally, building on Sat2Sound's ability to retrieve detailed soundscape captions, we introduce a novel application: location-based soundscape synthesis, which enables immersive acoustic experiences. Our code and models will be publicly available.
Knowledge graphs (KGs) have become an effective paradigm for managing real-world facts, which are not only complex but also dynamically evolve over time. The temporal validity of facts often serves as a strong clue in downstream link prediction tasks, which predicts a missing element in a fact. Traditional link prediction techniques on temporal KGs either consider a sequence of temporal snapshots of KGs with an ad-hoc defined time interval or expand a temporal fact over its validity period under a predefined time granularity; these approaches not only suffer from the sensitivity of the selection of time interval/granularity, but also face the computational challenges when handling facts with long (even infinite) validity. Although the recent hyper-relational KGs represent the temporal validity of a fact as qualifiers describing the fact, it is still suboptimal due to its ignorance of the infinite validity of some facts and the insufficient information encoded from the qualifiers about the temporal validity. Against this background, we propose VITA, a $\underline{V}$ersatile t$\underline{I}$me represen$\underline{TA}$tion learning method for temporal hyper-relational knowledge graphs. We first propose a versatile time representation that can flexibly accommodate all four types of temporal validity of facts (i.e., since, until, period, time-invariant), and then design VITA to effectively learn the time information in both aspects of time value and timespan to boost the link prediction performance. We conduct a thorough evaluation of VITA compared to a sizable collection of baselines on real-world KG datasets. Results show that VITA outperforms the best-performing baselines in various link prediction tasks (predicting missing entities, relations, time, and other numeric literals) by up to 75.3%. Ablation studies and a case study also support our key design choices.
Training vision-based manipulation policies that are robust across diverse visual environments remains an important and unresolved challenge in robot learning. Current approaches often sidestep the problem by relying on invariant representations such as point clouds and depth, or by brute-forcing generalization through visual domain randomization and/or large, visually diverse datasets. Disentangled representation learning - especially when combined with principles of associative memory - has recently shown promise in enabling vision-based reinforcement learning policies to be robust to visual distribution shifts. However, these techniques have largely been constrained to simpler benchmarks and toy environments. In this work, we scale disentangled representation learning and associative memory to more visually and dynamically complex manipulation tasks and demonstrate zero-shot adaptability to visual perturbations in both simulation and on real hardware. We further extend this approach to imitation learning, specifically Diffusion Policy, and empirically show significant gains in visual generalization compared to state-of-the-art imitation learning methods. Finally, we introduce a novel technique adapted from the model equivariance literature that transforms any trained neural network policy into one invariant to 2D planar rotations, making our policy not only visually robust but also resilient to certain camera perturbations. We believe that this work marks a significant step towards manipulation policies that are not only adaptable out of the box, but also robust to the complexities and dynamical nature of real-world deployment. Supplementary videos are available at https://sites.google.com/view/vis-gen-robotics/home.
Existing point cloud representation learning tend to learning the geometric distribution of objects through data-driven approaches, emphasizing structural features while overlooking the relationship between the local information and the whole structure. Local features reflect the fine-grained variations of an object, while the whole structure is determined by the interaction and combination of these local features, collectively defining the object's shape. In real-world, objects undergo elastic deformation under external forces, and this deformation gradually affects the whole structure through the propagation of forces from local regions, thereby altering the object's geometric properties. Inspired by this, we propose a physics-driven self-supervised learning method for point cloud representation, which captures the relationship between parts and the whole by constructing a local-whole force propagation mechanism. Specifically, we employ a dual-task encoder-decoder framework, integrating the geometric modeling capability of implicit fields with physics-driven elastic deformation. The encoder extracts features from the point cloud and its tetrahedral mesh representation, capturing both geometric and physical properties. These features are then fed into two decoders: one learns the whole geometric shape of the point cloud through an implicit field, while the other predicts local deformations using two specifically designed physics information loss functions, modeling the deformation relationship between local and whole shapes. Experimental results show that our method outperforms existing approaches in object classification, few-shot learning, and segmentation, demonstrating its effectiveness.
Graph neural networks (GNNs) have advanced recommender systems by modeling interaction relationships. However, existing graph-based recommenders rely on sparse ID features and do not fully exploit textual information, resulting in low information density within representations. Furthermore, graph contrastive learning faces challenges. Random negative sampling can introduce false negative samples, while fixed temperature coefficients cannot adapt to the heterogeneity of different nodes. In addition, current efforts to enhance recommendations with large language models (LLMs) have not fully utilized their Chain-of-Thought (CoT) reasoning capabilities to guide representation learning. To address these limitations, we introduces LGHRec (LLM-CoT Enhanced Graph Neural Recommendation with Harmonized Group Policy Optimization). This framework leverages the CoT reasoning ability of LLMs to generate semantic IDs, enriching reasoning processes and improving information density and semantic quality of representations. Moreover, we design a reinforcement learning algorithm, Harmonized Group Policy Optimization (HGPO), to optimize negative sampling strategies and temperature coefficients in contrastive learning. This approach enhances long-tail recommendation performance and ensures optimization consistency across different groups. Experimental results on three datasets demonstrate that LGHRec improves representation quality through semantic IDs generated by LLM's CoT reasoning and effectively boosts contrastive learning with HGPO. Our method outperforms several baseline models. The code is available at: https://anonymous.4open.science/r/LLM-Rec.
This work investigates the use of Wasserstein correlation -- a normalized measure of statistical dependence based on the Wasserstein distance between a joint distribution and the product of its marginals -- for unsupervised representation learning. Unlike, for example, contrastive methods, which naturally cluster classes in the latent space, we find that an (auto)encoder trained to maximize Wasserstein correlation between the input and encoded distributions instead acts as a compressor, reducing dimensionality while approximately preserving the topological and geometric properties of the input distribution. More strikingly, we show that Wasserstein correlation maximization can be used to arrive at an (auto)encoder -- either trained from scratch, or else one that extends a frozen, pretrained model -- that is approximately invariant to a chosen augmentation, or collection of augmentations, and that still approximately preserves the structural properties of the non-augmented input distribution. To do this, we first define the notion of an augmented encoder using the machinery of Markov-Wasserstein kernels. When the maximization objective is then applied to the augmented encoder, as opposed to the underlying, deterministic encoder, the resulting model exhibits the desired invariance properties. Finally, besides our experimental results, which show that even simple feedforward networks can be imbued with invariants or can, alternatively, be used to impart invariants to pretrained models under this training process, we additionally establish various theoretical results for optimal transport-based dependence measures. Code is available at https://github.com/keenan-eikenberry/wasserstein_correlation_maximization .