Graph-based multi-view spectral clustering methods have achieved notable progress recently, yet they often fall short in either oversimplifying pairwise relationships or struggling with inefficient spectral decompositions in high-dimensional Euclidean spaces. In this paper, we introduce a novel approach that begins to generate hypergraphs by leveraging sparse representation learning from data points. Based on the generated hypergraph, we propose an optimization function with orthogonality constraints for multi-view hypergraph spectral clustering, which incorporates spectral clustering for each view and ensures consistency across different views. In Euclidean space, solving the orthogonality-constrained optimization problem may yield local maxima and approximation errors. Innovately, we transform this problem into an unconstrained form on the Grassmannian manifold. Finally, we devise an alternating iterative Riemannian optimization algorithm to solve the problem. To validate the effectiveness of the proposed algorithm, we test it on four real-world multi-view datasets and compare its performance with seven state-of-the-art multi-view clustering algorithms. The experimental results demonstrate that our method outperforms the baselines in terms of clustering performance due to its superior low-dimensional and resilient feature representation.
Audio-Visual Speech Recognition (AVSR) leverages both audio and visual modalities to enhance speech recognition robustness, particularly in noisy environments. Recent advancements in Large Language Models (LLMs) have demonstrated their effectiveness in speech recognition, including AVSR. However, due to the significant length of speech representations, direct integration with LLMs imposes substantial computational costs. Prior approaches address this by compressing speech representations before feeding them into LLMs. However, higher compression ratios often lead to performance degradation, necessitating a trade-off between computational efficiency and recognition accuracy. To address this challenge, we propose Llama-MTSK, the first Matryoshka-based Multimodal LLM for AVSR, which enables flexible adaptation of the audio-visual token allocation based on specific computational constraints while preserving high performance. Our approach, inspired by Matryoshka Representation Learning, encodes audio-visual representations at multiple granularities within a single model, eliminating the need to train separate models for different compression levels. Moreover, to efficiently fine-tune the LLM, we introduce three LoRA-based Matryoshka strategies using global and scale-specific LoRA modules. Extensive evaluations on the two largest AVSR datasets demonstrate that Llama-MTSK achieves state-of-the-art results, matching or surpassing models trained independently at fixed compression levels.
Reinforced random walks (RRWs), including vertex-reinforced random walks (VRRWs) and edge-reinforced random walks (ERRWs), model random walks where the transition probabilities evolve based on prior visitation history~\cite{mgr, fmk, tarres, volkov}. These models have found applications in various areas, such as network representation learning~\cite{xzzs}, reinforced PageRank~\cite{gly}, and modeling animal behaviors~\cite{smouse}, among others. However, statistical estimation of the parameters governing RRWs remains underexplored. This work focuses on estimating the initial edge weights of ERRWs using observed trajectory data. Leveraging the connections between an ERRW and a random walk in a random environment (RWRE)~\cite{mr, mr2}, as given by the so-called "magic formula", we propose an estimator based on the generalized method of moments. To analyze the sample complexity of our estimator, we exploit the hyperbolic Gaussian structure embedded in the random environment to bound the fluctuations of the underlying random edge conductances.
This paper proposes a new principled multi-task representation learning framework (InfoMTL) to extract noise-invariant sufficient representations for all tasks. It ensures sufficiency of shared representations for all tasks and mitigates the negative effect of redundant features, which can enhance language understanding of pre-trained language models (PLMs) under the multi-task paradigm. Firstly, a shared information maximization principle is proposed to learn more sufficient shared representations for all target tasks. It can avoid the insufficiency issue arising from representation compression in the multi-task paradigm. Secondly, a task-specific information minimization principle is designed to mitigate the negative effect of potential redundant features in the input for each task. It can compress task-irrelevant redundant information and preserve necessary information relevant to the target for multi-task prediction. Experiments on six classification benchmarks show that our method outperforms 12 comparative multi-task methods under the same multi-task settings, especially in data-constrained and noisy scenarios. Extensive experiments demonstrate that the learned representations are more sufficient, data-efficient, and robust.
Existing methods for self-supervised representation learning of geospatial regions and map entities rely extensively on the design of pretext tasks, often involving augmentations or heuristic sampling of positive and negative pairs based on spatial proximity. This reliance introduces biases and limits the representations' expressiveness and generalisability. Consequently, the literature has expressed a pressing need to explore different methods for modelling geospatial data. To address the key difficulties of such methods, namely multimodality, heterogeneity, and the choice of pretext tasks, we present GeoJEPA, a versatile multimodal fusion model for geospatial data built on the self-supervised Joint-Embedding Predictive Architecture. With GeoJEPA, we aim to eliminate the widely accepted augmentation- and sampling biases found in self-supervised geospatial representation learning. GeoJEPA uses self-supervised pretraining on a large dataset of OpenStreetMap attributes, geometries and aerial images. The results are multimodal semantic representations of urban regions and map entities that we evaluate both quantitatively and qualitatively. Through this work, we uncover several key insights into JEPA's ability to handle multimodal data.
Multimodal Large Language Models (MLLMs) demonstrate robust zero-shot capabilities across diverse vision-language tasks after training on mega-scale datasets. However, dense prediction tasks, such as semantic segmentation and keypoint detection, pose significant challenges for MLLMs when represented solely as text outputs. Simultaneously, current MLLMs utilizing latent embeddings for visual task decoding generally demonstrate limited adaptability to both multi-task learning and multi-granularity scenarios. In this work, we present REF-VLM, an end-to-end framework for unified training of various visual decoding tasks. To address complex visual decoding scenarios, we introduce the Triplet-Based Referring Paradigm (TRP), which explicitly decouples three critical dimensions in visual decoding tasks through a triplet structure: concepts, decoding types, and targets. TRP employs symbolic delimiters to enforce structured representation learning, enhancing the parsability and interpretability of model outputs. Additionally, we construct Visual-Task Instruction Following Dataset (VTInstruct), a large-scale multi-task dataset containing over 100 million multimodal dialogue samples across 25 task types. Beyond text inputs and outputs, VT-Instruct incorporates various visual prompts such as point, box, scribble, and mask, and generates outputs composed of text and visual units like box, keypoint, depth and mask. The combination of different visual prompts and visual units generates a wide variety of task types, expanding the applicability of REF-VLM significantly. Both qualitative and quantitative experiments demonstrate that our REF-VLM outperforms other MLLMs across a variety of standard benchmarks. The code, dataset, and demo available at https://github.com/MacavityT/REF-VLM.
In oncology, Positron Emission Tomography-Computed Tomography (PET/CT) is widely used in cancer diagnosis, staging, and treatment monitoring, as it combines anatomical details from CT with functional metabolic activity and molecular marker expression information from PET. However, existing artificial intelligence-driven PET/CT analyses rely predominantly on task-specific models trained from scratch or on limited datasets, limiting their generalizability and robustness. To address this, we propose a foundation model approach specifically designed for multimodal PET/CT imaging. We introduce the Cross-Fraternal Twin Masked Autoencoder (FratMAE), a novel framework that effectively integrates whole-body anatomical and functional or molecular information. FratMAE employs separate Vision Transformer (ViT) encoders for PET and CT scans, along with cross-attention decoders that enable synergistic interactions between modalities during masked autoencoder training. Additionally, it incorporates textual metadata to enhance PET representation learning. By pre-training on PET/CT datasets, FratMAE captures intricate cross-modal relationships and global uptake patterns, achieving superior performance on downstream tasks and demonstrating its potential as a generalizable foundation model.
A major challenge in causal discovery from observational data is the absence of perfect interventions, making it difficult to distinguish causal features from spurious ones. We propose an innovative approach, Feature Matching Intervention (FMI), which uses a matching procedure to mimic perfect interventions. We define causal latent graphs, extending structural causal models to latent feature space, providing a framework that connects FMI with causal graph learning. Our feature matching procedure emulates perfect interventions within these causal latent graphs. Theoretical results demonstrate that FMI exhibits strong out-of-distribution (OOD) generalizability. Experiments further highlight FMI's superior performance in effectively identifying causal features solely from observational data.
Deep Learning has emerged as one of the most significant innovations in machine learning. However, a notable limitation of this field lies in the ``black box" decision-making processes, which have led to skepticism within groups like healthcare and scientific communities regarding its applicability. In response, this study introduces a interpretable approach using Neural Ordinary Differential Equations (NODEs), a category of neural network models that exploit the dynamics of differential equations for representation learning. Leveraging their foundation in differential equations, we illustrate the capability of these models to continuously process textual data, marking the first such model of its kind, and thereby proposing a promising direction for future research in this domain. The primary objective of this research is to propose a novel architecture for groups like healthcare that require the predictive capabilities of deep learning while emphasizing the importance of model transparency demonstrated in NODEs.
Stability for dynamic network embeddings ensures that nodes behaving the same at different times receive the same embedding, allowing comparison of nodes in the network across time. We present attributed unfolded adjacency spectral embedding (AUASE), a stable unsupervised representation learning framework for dynamic networks in which nodes are attributed with time-varying covariate information. To establish stability, we prove uniform convergence to an associated latent position model. We quantify the benefits of our dynamic embedding by comparing with state-of-the-art network representation learning methods on three real attributed networks. To the best of our knowledge, AUASE is the only attributed dynamic embedding that satisfies stability guarantees without the need for ground truth labels, which we demonstrate provides significant improvements for link prediction and node classification.
Deep learning scaling laws predict how performance improves with increased model and dataset size. Here we identify measurement noise in data as another performance scaling axis, governed by a distinct logarithmic law. We focus on representation learning models of biological single cell genomic data, where a dominant source of measurement noise is due to molecular undersampling. We introduce an information-theoretic metric for cellular representation model quality, and find that it scales with sampling depth. A single quantitative relationship holds across several model types and across several datasets. We show that the analytical form of this relationship can be derived from a simple Gaussian noise model, which in turn provides an intuitive interpretation for the scaling law. Finally, we show that the same relationship emerges in image classification models with respect to two types of imaging noise, suggesting that measurement noise scaling may be a general phenomenon. Scaling with noise can serve as a guide in generating and curating data for deep learning models, particularly in fields where measurement quality can vary dramatically between datasets.
In this study, we implement a Wasserstein GAN with Gradient Penalty (WGAN-GP) to generate multi-channel resting-state EEG data and assess the quality of the synthesized signals through both visual and feature-based evaluations. Our results indicate that the model effectively captures the statistical and spectral characteristics of real EEG data, although challenges remain in replicating high-frequency oscillations in the frontal region. Additionally, we demonstrate that the Critic's learned representations can be reused for gender classification task, achieving an out-of-sample accuracy, significantly better than a shuffled-label baseline and a model trained directly on EEG data. These findings suggest that generative models can serve not only as EEG data generators but also as unsupervised feature extractors, reducing the need for manual feature engineering. This study highlights the potential of GAN-based unsupervised learning for EEG analysis, suggesting avenues for more data-efficient deep learning applications in neuroscience.
Understanding how information is represented in neural networks is a fundamental challenge in both neuroscience and artificial intelligence. Despite their nonlinear architectures, recent evidence suggests that neural networks encode features in superposition, meaning that input concepts are linearly overlaid within the network's representations. We present a perspective that explains this phenomenon and provides a foundation for extracting interpretable representations from neural activations. Our theoretical framework consists of three steps: (1) Identifiability theory shows that neural networks trained for classification recover latent features up to a linear transformation. (2) Sparse coding methods can extract disentangled features from these representations by leveraging principles from compressed sensing. (3) Quantitative interpretability metrics provide a means to assess the success of these methods, ensuring that extracted features align with human-interpretable concepts. By bridging insights from theoretical neuroscience, representation learning, and interpretability research, we propose an emerging perspective on understanding neural representations in both artificial and biological systems. Our arguments have implications for neural coding theories, AI transparency, and the broader goal of making deep learning models more interpretable.
In representation learning, uniformity refers to the uniform feature distribution in the latent space (i.e., unit hypersphere). Previous work has shown that improving uniformity contributes to the learning of under-represented classes. However, most of the previous work focused on classification; the representation space of imbalanced regression remains unexplored. Classification-based methods are not suitable for regression tasks because they cluster features into distinct groups without considering the continuous and ordered nature essential for regression. In a geometric aspect, we uniquely focus on ensuring uniformity in the latent space for imbalanced regression through two key losses: enveloping and homogeneity. The enveloping loss encourages the induced trace to uniformly occupy the surface of a hypersphere, while the homogeneity loss ensures smoothness, with representations evenly spaced at consistent intervals. Our method integrates these geometric principles into the data representations via a Surrogate-driven Representation Learning (SRL) framework. Experiments with real-world regression and operator learning tasks highlight the importance of uniformity in imbalanced regression and validate the efficacy of our geometry-based loss functions.
Multimodal representation learning has demonstrated remarkable potential in enabling models to process and integrate diverse data modalities, such as text and images, for improved understanding and performance. While the medical domain can benefit significantly from this paradigm, the scarcity of paired multimodal data and reliance on proprietary or pretrained encoders pose significant challenges. In this work, we present a shared encoder framework for multimodal representation learning tailored to the medical domain. Our approach employs a single set of encoder parameters shared across modalities, augmented with learnable modality features. Empirical results demonstrate that our shared encoder idea achieves superior performance compared to separate modality-specific encoders, demonstrating improved generalization in data-constrained settings. Notably, the performance gains are more pronounced with fewer training examples, underscoring the efficiency of our shared encoder framework for real-world medical applications with limited data. Our code and experiment setup are available at https://github.com/VectorInstitute/shared_encoder.
Graph Neural Networks (GNNs) set the state-of-the-art in representation learning for graph-structured data. They are used in many domains, from online social networks to complex molecules. Most GNNs leverage the message-passing paradigm and achieve strong performances on various tasks. However, the message-passing mechanism used in most models suffers from over-smoothing as a GNN's depth increases. The over-smoothing degrades GNN's performance due to the increased similarity between the representations of unrelated nodes. This study proposes an adaptive channel-wise message-passing approach to alleviate the over-smoothing. The proposed model, Channel-Attentive GNN, learns how to attend to neighboring nodes and their feature channels. Thus, much diverse information can be transferred between nodes during message-passing. Experiments with widely used benchmark datasets show that the proposed model is more resistant to over-smoothing than baselines and achieves state-of-the-art performances for various graphs with strong heterophily. Our code is at https://github.com/ALLab-Boun/CHAT-GNN.
Multitask Representation Learning (MRL) has emerged as a prevalent technique to improve sample efficiency in Reinforcement Learning (RL). Empirical studies have found that training agents on multiple tasks simultaneously within online and transfer learning environments can greatly improve efficiency. Despite its popularity, a comprehensive theoretical framework that elucidates its operational efficacy remains incomplete. Prior analyses have predominantly assumed that agents either possess a pre-known representation function or utilize functions from a linear class, where both are impractical. The complexity of real-world applications typically requires the use of sophisticated, non-linear functions such as neural networks as representation function, which are not pre-existing but must be learned. Our work tries to fill the gap by extending the analysis to \textit{unknown non-linear} representations, giving a comprehensive analysis for its mechanism in online and transfer learning setting. We consider the setting that an agent simultaneously playing $M$ contextual bandits (or MDPs), developing a shared representation function $\phi$ from a non-linear function class $\Phi$ using our novel Generalized Functional Upper Confidence Bound algorithm (GFUCB). We formally prove that this approach yields a regret upper bound that outperforms the lower bound associated with learning $M$ separate tasks, marking the first demonstration of MRL's efficacy in a general function class. This framework also explains the contribution of representations to transfer learning when faced with new, yet related tasks, and identifies key conditions for successful transfer. Empirical experiments further corroborate our theoretical findings.
In egocentric video understanding, the motion of hands and objects as well as their interactions play a significant role by nature. However, existing egocentric video representation learning methods mainly focus on aligning video representation with high-level narrations, overlooking the intricate dynamics between hands and objects. In this work, we aim to integrate the modeling of fine-grained hand-object dynamics into the video representation learning process. Since no suitable data is available, we introduce HOD, a novel pipeline employing a hand-object detector and a large language model to generate high-quality narrations with detailed descriptions of hand-object dynamics. To learn these fine-grained dynamics, we propose EgoVideo, a model with a new lightweight motion adapter to capture fine-grained hand-object motion information. Through our co-training strategy, EgoVideo effectively and efficiently leverages the fine-grained hand-object dynamics in the HOD data. Extensive experiments demonstrate that our method achieves state-of-the-art performance across multiple egocentric downstream tasks, including improvements of 6.3% in EK-100 multi-instance retrieval, 5.7% in EK-100 classification, and 16.3% in EGTEA classification in zero-shot settings. Furthermore, our model exhibits robust generalization capabilities in hand-object interaction and robot manipulation tasks. Code and data are available at https://github.com/OpenRobotLab/EgoHOD/.
Masked Image Modeling (MIM) has garnered significant attention in self-supervised learning, thanks to its impressive capacity to learn scalable visual representations tailored for downstream tasks. However, images inherently contain abundant redundant information, leading the pixel-based MIM reconstruction process to focus excessively on finer details such as textures, thus prolonging training times unnecessarily. Addressing this challenge requires a shift towards a compact representation of features during MIM reconstruction. Frequency domain analysis provides a promising avenue for achieving compact image feature representation. In contrast to the commonly used Fourier transform, wavelet transform not only offers frequency information but also preserves spatial characteristics and multi-level features of the image. Additionally, the multi-level decomposition process of wavelet transformation aligns well with the hierarchical architecture of modern neural networks. In this study, we leverage wavelet transform as a tool for efficient representation learning to expedite the training process of MIM. Specifically, we conduct multi-level decomposition of images using wavelet transform, utilizing wavelet coefficients from different levels to construct distinct reconstruction targets representing various frequencies and scales. These reconstruction targets are then integrated into the MIM process, with adjustable weights assigned to prioritize the most crucial information. Extensive experiments demonstrate that our method achieves comparable or superior performance across various downstream tasks while exhibiting higher training efficiency.
Pre-training and representation learning have been playing an increasingly important role in modern speech processing. Nevertheless, different applications have been relying on different foundation models, since predominant pre-training techniques are either designed for discriminative tasks or generative tasks. In this work, we make the first attempt at building a unified pre-training framework for both types of tasks in speech. We show that with the appropriate design choices for pre-training, one can jointly learn a representation encoder and generative audio decoder that can be applied to both types of tasks. We propose UniWav, an encoder-decoder framework designed to unify pre-training representation learning and generative tasks. On speech recognition, text-to-speech, and speech tokenization, UniWav achieves comparable performance to different existing foundation models, each trained on a specific task. Our findings suggest that a single general-purpose foundation model for speech can be built to replace different foundation models, reducing the overhead and cost of pre-training.