Self-supervised learning (SSL) approaches have brought tremendous success across many tasks and domains. It has been argued that these successes can be attributed to a link between SSL and identifiable representation learning: Temporal structure and auxiliary variables ensure that latent representations are related to the true underlying generative factors of the data. Here, we deepen this connection and show that SSL can perform system identification in latent space. We propose DynCL, a framework to uncover linear, switching linear and non-linear dynamics under a non-linear observation model, give theoretical guarantees and validate them empirically.
Self-supervised learning systems have gained significant attention in recent years by leveraging clustering-based pseudo-labels to provide supervision without the need for human annotations. However, the noise in these pseudo-labels caused by the clustering methods poses a challenge to the learning process leading to degraded performance. In this work, we propose a pseudo-label refinement (SLR) algorithm to address this issue. The cluster labels from the previous epoch are projected to the current epoch cluster-labels space and a linear combination of the new label and the projected label is computed as a soft refined label containing the information from the previous epoch clusters as well as from the current epoch. In contrast to the common practice of using the maximum value as a cluster/class indicator, we employ hierarchical clustering on these soft pseudo-labels to generate refined hard-labels. This approach better utilizes the information embedded in the soft labels, outperforming the simple maximum value approach for hard label generation. The effectiveness of the proposed SLR algorithm is evaluated in the context of person re-identification (Re-ID) using unsupervised domain adaptation (UDA). Experimental results demonstrate that the modified Re-ID baseline, incorporating the SLR algorithm, achieves significantly improved mean Average Precision (mAP) performance in various UDA tasks, including real-to-synthetic, synthetic-to-real, and different real-to-real scenarios. These findings highlight the efficacy of the SLR algorithm in enhancing the performance of self-supervised learning systems.
The development of 3D medical vision-language models holds significant potential for disease diagnosis and patient treatment. However, compared to 2D medical images, 3D medical images, such as CT scans, face challenges related to limited training data and high dimension, which severely restrict the progress of 3D medical vision-language models. To address these issues, we collect a large amount of unlabeled 3D CT data and utilize self-supervised learning to construct a 3D visual foundation model for extracting 3D visual features. Then, we apply 3D spatial convolutions to aggregate and project high-level image features, reducing computational complexity while preserving spatial information. We also construct two instruction-tuning datasets based on BIMCV-R and CT-RATE to fine-tune the 3D vision-language model. Our model demonstrates superior performance compared to existing methods in report generation, visual question answering, and disease diagnosis. Code and data will be made publicly available soon.
A prominent self-supervised learning paradigm is to model the representations as clusters, or more generally as a mixture model. Learning to map the data samples to compact representations and fitting the mixture model simultaneously leads to the representation collapse problem. Regularizing the distribution of data points over the clusters is the prevalent strategy to avoid this issue. While this is sufficient to prevent full representation collapse, we show that a partial prototype collapse problem still exists in the DINO family of methods, that leads to significant redundancies in the prototypes. Such prototype redundancies serve as shortcuts for the method to achieve a marginal latent class distribution that matches the prescribed prior. We show that by encouraging the model to use diverse prototypes, the partial prototype collapse can be mitigated. Effective utilization of the prototypes enables the methods to learn more fine-grained clusters, encouraging more informative representations. We demonstrate that this is especially beneficial when pre-training on a long-tailed fine-grained dataset.
Despite their impressive success, training foundation models remains computationally costly. This paper investigates how to efficiently train speech foundation models with self-supervised learning (SSL) under a limited compute budget. We examine critical factors in SSL that impact the budget, including model architecture, model size, and data size. Our goal is to make analytical steps toward understanding the training dynamics of speech foundation models. We benchmark SSL objectives in an entirely comparable setting and find that other factors contribute more significantly to the success of SSL. Our results show that slimmer model architectures outperform common small architectures under the same compute and parameter budget. We demonstrate that the size of the pre-training data remains crucial, even with data augmentation during SSL training, as performance suffers when iterating over limited data. Finally, we identify a trade-off between model size and data size, highlighting an optimal model size for a given compute budget.
We introduce MODEL SELECTOR, a framework for label-efficient selection of pretrained classifiers. Given a pool of unlabeled target data, MODEL SELECTOR samples a small subset of highly informative examples for labeling, in order to efficiently identify the best pretrained model for deployment on this target dataset. Through extensive experiments, we demonstrate that MODEL SELECTOR drastically reduces the need for labeled data while consistently picking the best or near-best performing model. Across 18 model collections on 16 different datasets, comprising over 1,500 pretrained models, MODEL SELECTOR reduces the labeling cost by up to 94.15% to identify the best model compared to the cost of the strongest baseline. Our results further highlight the robustness of MODEL SELECTOR in model selection, as it reduces the labeling cost by up to 72.41% when selecting a near-best model, whose accuracy is only within 1% of the best model.
Self-supervised learning on tabular data seeks to apply advances from natural language and image domains to the diverse domain of tables. However, current techniques often struggle with integrating multi-domain data and require data cleaning or specific structural requirements, limiting the scalability of pre-training datasets. We introduce PORTAL (Pretraining One-Row-at-a-Time for All tabLes), a framework that handles various data modalities without the need for cleaning or preprocessing. This simple yet powerful approach can be effectively pre-trained on online-collected datasets and fine-tuned to match state-of-the-art methods on complex classification and regression tasks. This work offers a practical advancement in self-supervised learning for large-scale tabular data.
Semantic segmentation of remote sensing (RS) images is a challenging yet essential task with broad applications. While deep learning, particularly supervised learning with large-scale labeled datasets, has significantly advanced this field, the acquisition of high-quality labeled data remains costly and time-intensive. Unsupervised domain adaptation (UDA) provides a promising alternative by enabling models to learn from unlabeled target domain data while leveraging labeled source domain data. Recent self-training (ST) approaches employing pseudo-label generation have shown potential in mitigating domain discrepancies. However, the application of ST to RS image segmentation remains underexplored. Factors such as variations in ground sampling distance, imaging equipment, and geographic diversity exacerbate domain shifts, limiting model performance across domains. In that case, existing ST methods, due to significant domain shifts in cross-domain RS images, often underperform. To address these challenges, we propose integrating contrastive learning into UDA, enhancing the model's ability to capture semantic information in the target domain by maximizing the similarity between augmented views of the same image. This additional supervision improves the model's representational capacity and segmentation performance in the target domain. Extensive experiments conducted on RS datasets, including Potsdam, Vaihingen, and LoveDA, demonstrate that our method, SimSeg, outperforms existing approaches, achieving state-of-the-art results. Visualization and quantitative analyses further validate SimSeg's superior ability to learn from the target domain. The code is publicly available at https://github.com/woldier/SiamSeg.
In this paper, we present EH-MAM (Easy-to-Hard adaptive Masked Acoustic Modeling), a novel self-supervised learning approach for speech representation learning. In contrast to the prior methods that use random masking schemes for Masked Acoustic Modeling (MAM), we introduce a novel selective and adaptive masking strategy. Specifically, during SSL training, we progressively introduce harder regions to the model for reconstruction. Our approach automatically selects hard regions and is built on the observation that the reconstruction loss of individual frames in MAM can provide natural signals to judge the difficulty of solving the MAM pre-text task for that frame. To identify these hard regions, we employ a teacher model that first predicts the frame-wise losses and then decides which frames to mask. By learning to create challenging problems, such as identifying harder frames and solving them simultaneously, the model is able to learn more effective representations and thereby acquire a more comprehensive understanding of the speech. Quantitatively, EH-MAM outperforms several state-of-the-art baselines across various low-resource speech recognition and SUPERB benchmarks by 5%-10%. Additionally, we conduct a thorough analysis to show that the regions masked by EH-MAM effectively capture useful context across speech frames.
Transformers serve as the backbone architectures of Foundational Models, where domain-specific tokenizers allow them to adapt to various domains. Graph Transformers (GTs) have recently emerged as leading models in geometric deep learning, outperforming Graph Neural Networks (GNNs) in various graph learning tasks. However, the development of tokenizers for graphs has lagged behind other modalities. To address this, we introduce GQT (\textbf{G}raph \textbf{Q}uantized \textbf{T}okenizer), which decouples tokenizer training from Transformer training by leveraging multi-task graph self-supervised learning, yielding robust and generalizable graph tokens. Furthermore, the GQT utilizes Residual Vector Quantization (RVQ) to learn hierarchical discrete tokens, resulting in significantly reduced memory requirements and improved generalization capabilities. By combining the GQT with token modulation, a Transformer encoder achieves state-of-the-art performance on 20 out of 22 benchmarks, including large-scale homophilic and heterophilic datasets.
Change point detection (CPD) methods aim to identify abrupt shifts in the distribution of input data streams. Accurate estimators for this task are crucial across various real-world scenarios. Yet, traditional unsupervised CPD techniques face significant limitations, often relying on strong assumptions or suffering from low expressive power due to inherent model simplicity. In contrast, representation learning methods overcome these drawbacks by offering flexibility and the ability to capture the full complexity of the data without imposing restrictive assumptions. However, these approaches are still emerging in the CPD field and lack robust theoretical foundations to ensure their reliability. Our work addresses this gap by integrating the expressive power of representation learning with the groundedness of traditional CPD techniques. We adopt spectral normalization (SN) for deep representation learning in CPD tasks and prove that the embeddings after SN are highly informative for CPD. Our method significantly outperforms current state-of-the-art methods during the comprehensive evaluation via three standard CPD datasets.
Pre-training foundation models has become the de-facto procedure for deep learning approaches, yet its application remains limited in the geological studies, where in needs of the model transferability to break the shackle of data scarcity. Here we target on the X-ray fluorescence (XRF) scanning data, a standard high-resolution measurement in extensive scientific drilling projects. We propose a scalable self-supervised learner, masked autoencoders on XRF spectra (MAX), to pre-train a foundation model covering geological records from multiple regions of the Pacific and Southern Ocean. In pre-training, we find that masking a high proportion of the input spectrum (50\%) yields a nontrivial and meaningful self-supervisory task. For downstream tasks, we select the quantification of XRF spectra into two costly geochemical measurements, CaCO$_3$ and total organic carbon, due to their importance in understanding the paleo-oceanic carbon system. Our results show that MAX, requiring only one-third of the data, outperforms models without pre-training in terms of quantification accuracy. Additionally, the model's generalizability improves by more than 60\% in zero-shot tests on new materials, with explainability further ensuring its robustness. Thus, our approach offers a promising pathway to overcome data scarcity in geological discovery by leveraging the self-supervised foundation model and fast-acquired XRF scanning data.
We present MultiCamCows2024, a farm-scale image dataset filmed across multiple cameras for the biometric identification of individual Holstein-Friesian cattle exploiting their unique black and white coat-patterns. Captured by three ceiling-mounted visual sensors covering adjacent barn areas over seven days on a working dairy farm, the dataset comprises 101, 329 images of 90 cows, plus the underlying original CCTV footage. The dataset is provided alongside full computer vision recognition baselines, that is both a supervised and self-supervised learning framework for individual cow identification trained on cattle tracklets. We report a performance above 96% single image identification accuracy from the dataset and demonstrate that combining data from multiple cameras during learning enhances self-supervised identification. We show that our framework enables fully automatic cattle identification, barring only the simple human verification of tracklet integrity during data collection. Crucially, our study highlights that multi-camera, supervised and self-supervised components in tandem not only deliver highly accurate individual cow identification but also achieve this efficiently with no labelling of cattle identities by humans at all. We argue that this improvement in efficacy has practical implications for livestock management, behaviour analysis, and agricultural monitoring. For full reproducibility and practical ease of use, we publish all key software and code including re-identification components and the species detector with this paper.
Image fusion is famous as an alternative solution to generate one high-quality image from multiple images in addition to image restoration from a single degraded image. The essence of image fusion is to integrate complementary information from source images. Existing fusion methods struggle with generalization across various tasks and often require labor-intensive designs, in which it is difficult to identify and extract useful information from source images due to the diverse requirements of each fusion task. Additionally, these methods develop highly specialized features for different downstream applications, hindering the adaptation to new and diverse downstream tasks. To address these limitations, we introduce DeFusion++, a novel framework that leverages self-supervised learning (SSL) to enhance the versatility of feature representation for different image fusion tasks. DeFusion++ captures the image fusion task-friendly representations from large-scale data in a self-supervised way, overcoming the constraints of limited fusion datasets. Specifically, we introduce two innovative pretext tasks: common and unique decomposition (CUD) and masked feature modeling (MFM). CUD decomposes source images into abstract common and unique components, while MFM refines these components into robust fused features. Jointly training of these tasks enables DeFusion++ to produce adaptable representations that can effectively extract useful information from various source images, regardless of the fusion task. The resulting fused representations are also highly adaptable for a wide range of downstream tasks, including image segmentation and object detection. DeFusion++ stands out by producing versatile fused representations that can enhance both the quality of image fusion and the effectiveness of downstream high-level vision tasks, simplifying the process with the elegant fusion framework.
Multivariate time-series data in fields like healthcare and industry are informative but challenging due to high dimensionality and lack of labels. Recent self-supervised learning methods excel in learning rich representations without labels but struggle with disentangled embeddings and inductive bias issues like transformation-invariance. To address these challenges, we introduce TimeDRL, a framework for multivariate time-series representation learning with dual-level disentangled embeddings. TimeDRL features: (i) disentangled timestamp-level and instance-level embeddings using a [CLS] token strategy; (ii) timestamp-predictive and instance-contrastive tasks for representation learning; and (iii) avoidance of augmentation methods to eliminate inductive biases. Experiments on forecasting and classification datasets show TimeDRL outperforms existing methods, with further validation in semi-supervised settings with limited labeled data.
Latent-based image generative models, such as Latent Diffusion Models (LDMs) and Mask Image Models (MIMs), have achieved notable success in image generation tasks. These models typically leverage reconstructive autoencoders like VQGAN or VAE to encode pixels into a more compact latent space and learn the data distribution in the latent space instead of directly from pixels. However, this practice raises a pertinent question: Is it truly the optimal choice? In response, we begin with an intriguing observation: despite sharing the same latent space, autoregressive models significantly lag behind LDMs and MIMs in image generation. This finding contrasts sharply with the field of NLP, where the autoregressive model GPT has established a commanding presence. To address this discrepancy, we introduce a unified perspective on the relationship between latent space and generative models, emphasizing the stability of latent space in image generative modeling. Furthermore, we propose a simple but effective discrete image tokenizer to stabilize the latent space for image generative modeling by applying K-Means on the latent features of self-supervised learning models. Experimental results show that image autoregressive modeling with our tokenizer (DiGIT) benefits both image understanding and image generation with the next token prediction principle, which is inherently straightforward for GPT models but challenging for other generative models. Remarkably, for the first time, a GPT-style autoregressive model for images outperforms LDMs, which also exhibits substantial improvement akin to GPT when scaling up model size. Our findings underscore the potential of an optimized latent space and the integration of discrete tokenization in advancing the capabilities of image generative models. The code is available at \url{https://github.com/DAMO-NLP-SG/DiGIT}.
Speech Emotion Recognition (SER) analyzes human emotions expressed through speech. Self-supervised learning (SSL) offers a promising approach to SER by learning meaningful representations from a large amount of unlabeled audio data. However, existing SSL-based methods rely on Global Average Pooling (GAP) to represent audio signals, treating speech and non-speech segments equally. This can lead to dilution of informative speech features by irrelevant non-speech information. To address this, the paper proposes Segmental Average Pooling (SAP), which selectively focuses on informative speech segments while ignoring non-speech segments. By applying both GAP and SAP to SSL features, our approach utilizes overall speech signal information from GAP and specific information from SAP, leading to improved SER performance. Experiments show state-of-the-art results on the IEMOCAP for English and superior performance on KEMDy19 for Korean datasets in both unweighted and weighted accuracies.
In the era of big data and Artificial Intelligence, an emerging paradigm is to utilize contrastive self-supervised learning to model large-scale heterogeneous data. Many existing foundation models benefit from the generalization capability of contrastive self-supervised learning by learning compact and high-quality representations without relying on any label information. Amidst the explosive advancements in foundation models across multiple domains, including natural language processing and computer vision, a thorough survey on heterogeneous contrastive learning for the foundation model is urgently needed. In response, this survey critically evaluates the current landscape of heterogeneous contrastive learning for foundation models, highlighting the open challenges and future trends of contrastive learning. In particular, we first present how the recent advanced contrastive learning-based methods deal with view heterogeneity and how contrastive learning is applied to train and fine-tune the multi-view foundation models. Then, we move to contrastive learning methods for task heterogeneity, including pretraining tasks and downstream tasks, and show how different tasks are combined with contrastive learning loss for different purposes. Finally, we conclude this survey by discussing the open challenges and shedding light on the future directions of contrastive learning.
Effective spatio-temporal prediction frameworks play a crucial role in urban sensing applications, including traffic analysis, human mobility behavior modeling, and citywide crime prediction. However, the presence of data noise and label sparsity in spatio-temporal data presents significant challenges for existing neural network models in learning effective and robust region representations. To address these challenges, we propose a novel spatio-temporal graph masked autoencoder paradigm that explores generative self-supervised learning for effective spatio-temporal data augmentation. Our proposed framework introduces a spatial-temporal heterogeneous graph neural encoder that captures region-wise dependencies from heterogeneous data sources, enabling the modeling of diverse spatial dependencies. In our spatio-temporal self-supervised learning paradigm, we incorporate a masked autoencoding mechanism on node representations and structures. This mechanism automatically distills heterogeneous spatio-temporal dependencies across regions over time, enhancing the learning process of dynamic region-wise spatial correlations. To validate the effectiveness of our STGMAE framework, we conduct extensive experiments on various spatio-temporal mining tasks. We compare our approach against state-of-the-art baselines. The results of these evaluations demonstrate the superiority of our proposed framework in terms of performance and its ability to address the challenges of spatial and temporal data noise and sparsity in practical urban sensing scenarios.
Universal Domain Adaptation (UniDA) aims to transfer knowledge from a labeled source domain to an unlabeled target domain without assuming how much the label-sets of the two domains intersect. The goal of UniDA is to achieve robust performance on the target domain across different intersection levels. However, existing literature has not sufficiently explored performance under extreme intersection levels. Our experiments reveal that state-of-the-art methods struggle when the source domain has significantly more non-overlapping classes than overlapping ones, a setting we refer to as Extreme UniDA. In this paper, we demonstrate that classical partial domain alignment, which focuses on aligning only overlapping-class data between domains, is limited in mitigating the bias of feature extractors toward source-private classes in extreme UniDA scenarios. We argue that feature extractors trained with source supervised loss distort the intrinsic structure of the target data due to the inherent differences between source-private classes and the target data. To mitigate this bias, we propose using self-supervised learning to preserve the structure of the target data. Our approach can be easily integrated into existing frameworks. We apply the proposed approach to two distinct training paradigms-adversarial-based and optimal-transport-based-and show consistent improvements across various intersection levels, with significant gains in extreme UniDA settings.