Recent successes in self-supervised learning (SSL) model spatial co-occurrences of visual features either by masking portions of an image or by aggressively cropping it. Here, we propose a new way to model spatial co-occurrences by aligning local representations (before pooling) with a global image representation. We present CO-SSL, a family of instance discrimination methods and show that it outperforms previous methods on several datasets, including ImageNet-1K where it achieves 71.5% of Top-1 accuracy with 100 pre-training epochs. CO-SSL is also more robust to noise corruption, internal corruption, small adversarial attacks, and large training crop sizes. Our analysis further indicates that CO-SSL learns highly redundant local representations, which offers an explanation for its robustness. Overall, our work suggests that aligning local and global representations may be a powerful principle of unsupervised category learning.
Recent self-supervised learning (SSL) models trained on human-like egocentric visual inputs substantially underperform on image recognition tasks compared to humans. These models train on raw, uniform visual inputs collected from head-mounted cameras. This is different from humans, as the anatomical structure of the retina and visual cortex relatively amplifies the central visual information, i.e. around humans' gaze location. This selective amplification in humans likely aids in forming object-centered visual representations. Here, we investigate whether focusing on central visual information boosts egocentric visual object learning. We simulate 5-months of egocentric visual experience using the large-scale Ego4D dataset and generate gaze locations with a human gaze prediction model. To account for the importance of central vision in humans, we crop the visual area around the gaze location. Finally, we train a time-based SSL model on these modified inputs. Our experiments demonstrate that focusing on central vision leads to better object-centered representations. Our analysis shows that the SSL model leverages the temporal dynamics of the gaze movements to build stronger visual representations. Overall, our work marks a significant step toward bio-inspired learning of visual representations.
Clouds play a key role in Earth's radiation balance with complex effects that introduce large uncertainties into climate models. Real-time 3D cloud data is essential for improving climate predictions. This study leverages geostationary imagery from MSG/SEVIRI and radar reflectivity measurements of cloud profiles from CloudSat/CPR to reconstruct 3D cloud structures. We first apply self-supervised learning (SSL) methods-Masked Autoencoders (MAE) and geospatially-aware SatMAE on unlabelled MSG images, and then fine-tune our models on matched image-profile pairs. Our approach outperforms state-of-the-art methods like U-Nets, and our geospatial encoding further improves prediction results, demonstrating the potential of SSL for cloud reconstruction.
Accurate imputation of missing laboratory values in electronic health records (EHRs) is critical to enable robust clinical predictions and reduce biases in AI systems in healthcare. Existing methods, such as variational autoencoders (VAEs) and decision tree-based approaches such as XGBoost, struggle to model the complex temporal and contextual dependencies in EHR data, mainly in underrepresented groups. In this work, we propose Lab-MAE, a novel transformer-based masked autoencoder framework that leverages self-supervised learning for the imputation of continuous sequential lab values. Lab-MAE introduces a structured encoding scheme that jointly models laboratory test values and their corresponding timestamps, enabling explicit capturing temporal dependencies. Empirical evaluation on the MIMIC-IV dataset demonstrates that Lab-MAE significantly outperforms the state-of-the-art baselines such as XGBoost across multiple metrics, including root mean square error (RMSE), R-squared (R2), and Wasserstein distance (WD). Notably, Lab-MAE achieves equitable performance across demographic groups of patients, advancing fairness in clinical predictions. We further investigate the role of follow-up laboratory values as potential shortcut features, revealing Lab-MAE's robustness in scenarios where such data is unavailable. The findings suggest that our transformer-based architecture, adapted to the characteristics of the EHR data, offers a foundation model for more accurate and fair clinical imputation models. In addition, we measure and compare the carbon footprint of Lab-MAE with the baseline XGBoost model, highlighting its environmental requirements.
Contrastive learning, a prominent approach within self-supervised learning, has demonstrated significant effectiveness in developing generalizable models for various applications involving natural images. However, recent research indicates that these successes do not necessarily extend to the medical imaging domain. In this paper, we investigate the reasons for this suboptimal performance and hypothesize that the dense distribution of medical images poses challenges to the pretext tasks in contrastive learning, particularly in constructing positive and negative pairs. We explore model performance under different augmentation strategies and compare the results to those achieved with strong augmentations. Our study includes six publicly available datasets covering multiple clinically relevant tasks. We further assess the model's generalizability through external evaluations. The model pre-trained with weak augmentation outperforms those with strong augmentation, improving AUROC from 0.838 to 0.848 and AUPR from 0.523 to 0.597 on MESSIDOR2, and showing similar enhancements across other datasets. Our findings suggest that optimizing the scale of augmentation is critical for enhancing the efficacy of contrastive learning in medical imaging.
This paper explores Masked Autoencoders (MAE) with Gaussian Splatting. While reconstructive self-supervised learning frameworks such as MAE learns good semantic abstractions, it is not trained for explicit spatial awareness. Our approach, named Gaussian Masked Autoencoder, or GMAE, aims to learn semantic abstractions and spatial understanding jointly. Like MAE, it reconstructs the image end-to-end in the pixel space, but beyond MAE, it also introduces an intermediate, 3D Gaussian-based representation and renders images via splatting. We show that GMAE can enable various zero-shot learning capabilities of spatial understanding (e.g., figure-ground segmentation, image layering, edge detection, etc.) while preserving the high-level semantics of self-supervised representation quality from MAE. To our knowledge, we are the first to employ Gaussian primitives in an image representation learning framework beyond optimization-based single-scene reconstructions. We believe GMAE will inspire further research in this direction and contribute to developing next-generation techniques for modeling high-fidelity visual data. More details at https://brjathu.github.io/gmae
Image composition involves extracting a foreground object from one image and pasting it into another image through Image harmonization algorithms (IHAs), which aim to adjust the appearance of the foreground object to better match the background. Existing image quality assessment (IQA) methods may fail to align with human visual preference on image harmonization due to the insensitivity to minor color or light inconsistency. To address the issue and facilitate the advancement of IHAs, we introduce the first Image Quality Assessment Database for image Harmony evaluation (HarmonyIQAD), which consists of 1,350 harmonized images generated by 9 different IHAs, and the corresponding human visual preference scores. Based on this database, we propose a Harmony Image Quality Assessment (HarmonyIQA), to predict human visual preference for harmonized images. Extensive experiments show that HarmonyIQA achieves state-of-the-art performance on human visual preference evaluation for harmonized images, and also achieves competing results on traditional IQA tasks. Furthermore, cross-dataset evaluation also shows that HarmonyIQA exhibits better generalization ability than self-supervised learning-based IQA methods. Both HarmonyIQAD and HarmonyIQA will be made publicly available upon paper publication.
Recent years have witnessed the success of foundation models pre-trained with self-supervised learning (SSL) in various music informatics understanding tasks, including music tagging, instrument classification, key detection, and more. In this paper, we propose a self-supervised music representation learning model for music understanding. Distinguished from previous studies adopting random projection or existing neural codec, the proposed model, named MuQ, is trained to predict tokens generated by Mel Residual Vector Quantization (Mel-RVQ). Our Mel-RVQ utilizes residual linear projection structure for Mel spectrum quantization to enhance the stability and efficiency of target extraction and lead to better performance. Experiments in a large variety of downstream tasks demonstrate that MuQ outperforms previous self-supervised music representation models with only 0.9K hours of open-source pre-training data. Scaling up the data to over 160K hours and adopting iterative training consistently improve the model performance. To further validate the strength of our model, we present MuQ-MuLan, a joint music-text embedding model based on contrastive learning, which achieves state-of-the-art performance in the zero-shot music tagging task on the MagnaTagATune dataset. Code and checkpoints are open source in https://github.com/tencent-ailab/MuQ.
We present STORM, a spatio-temporal reconstruction model designed for reconstructing dynamic outdoor scenes from sparse observations. Existing dynamic reconstruction methods often rely on per-scene optimization, dense observations across space and time, and strong motion supervision, resulting in lengthy optimization times, limited generalization to novel views or scenes, and degenerated quality caused by noisy pseudo-labels for dynamics. To address these challenges, STORM leverages a data-driven Transformer architecture that directly infers dynamic 3D scene representations--parameterized by 3D Gaussians and their velocities--in a single forward pass. Our key design is to aggregate 3D Gaussians from all frames using self-supervised scene flows, transforming them to the target timestep to enable complete (i.e., "amodal") reconstructions from arbitrary viewpoints at any moment in time. As an emergent property, STORM automatically captures dynamic instances and generates high-quality masks using only reconstruction losses. Extensive experiments on public datasets show that STORM achieves precise dynamic scene reconstruction, surpassing state-of-the-art per-scene optimization methods (+4.3 to 6.6 PSNR) and existing feed-forward approaches (+2.1 to 4.7 PSNR) in dynamic regions. STORM reconstructs large-scale outdoor scenes in 200ms, supports real-time rendering, and outperforms competitors in scene flow estimation, improving 3D EPE by 0.422m and Acc5 by 28.02%. Beyond reconstruction, we showcase four additional applications of our model, illustrating the potential of self-supervised learning for broader dynamic scene understanding.
Speech Emotion Recognition (SER) involves analyzing vocal expressions to determine the emotional state of speakers, where the comprehensive and thorough utilization of audio information is paramount. Therefore, we propose a novel approach on self-supervised learning (SSL) models that employs all available auxiliary information -- specifically metadata -- to enhance performance. Through a two-stage fine-tuning method in multi-task learning, we introduce the Augmented Residual Integration (ARI) module, which enhances transformer layers in encoder of SSL models. The module efficiently preserves acoustic features across all different levels, thereby significantly improving the performance of metadata-related auxiliary tasks that require various levels of features. Moreover, the Co-attention module is incorporated due to its complementary nature with ARI, enabling the model to effectively utilize multidimensional information and contextual relationships from metadata-related auxiliary tasks. Under pre-trained base models and speaker-independent setup, our approach consistently surpasses state-of-the-art (SOTA) models on multiple SSL encoders for the IEMOCAP dataset.
The Emotional Voice Conversion (EVC) aims to convert the discrete emotional state from the source emotion to the target for a given speech utterance while preserving linguistic content. In this paper, we propose regularizing emotion intensity in the diffusion-based EVC framework to generate precise speech of the target emotion. Traditional approaches control the intensity of an emotional state in the utterance via emotion class probabilities or intensity labels that often lead to inept style manipulations and degradations in quality. On the contrary, we aim to regulate emotion intensity using self-supervised learning-based feature representations and unsupervised directional latent vector modeling (DVM) in the emotional embedding space within a diffusion-based framework. These emotion embeddings can be modified based on the given target emotion intensity and the corresponding direction vector. Furthermore, the updated embeddings can be fused in the reverse diffusion process to generate the speech with the desired emotion and intensity. In summary, this paper aims to achieve high-quality emotional intensity regularization in the diffusion-based EVC framework, which is the first of its kind work. The effectiveness of the proposed method has been shown across state-of-the-art (SOTA) baselines in terms of subjective and objective evaluations for the English and Hindi languages \footnote{Demo samples are available at the following URL: \url{https://nirmesh-sony.github.io/EmoReg/}}.
In the context of personalized federated learning, existing approaches train a global model to extract transferable representations, based on which any client could train personalized models with a limited number of data samples. Self-supervised learning is considered a promising direction as the global model it produces is generic and facilitates personalization for all clients fairly. However, when data is heterogeneous across clients, the global model trained using SSL is unable to learn high-quality personalized models. In this paper, we show that when the global model is trained with SSL without modifications, its produced representations have fuzzy class boundaries. As a result, personalized learning within each client produces models with low accuracy. In order to improve SSL towards better accuracy without sacrificing its advantage in fairness, we propose Calibre, a new personalized federated learning framework designed to calibrate SSL representations by maintaining a suitable balance between more generic and more client-specific representations. Calibre is designed based on theoretically-sound properties, and introduces (1) a client-specific prototype loss as an auxiliary training objective; and (2) an aggregation algorithm guided by such prototypes across clients. Our experimental results in an extensive array of non-i.i.d.~settings show that Calibre achieves state-of-the-art performance in terms of both mean accuracy and fairness across clients. Code repo: https://github.com/TL-System/plato/tree/main/examples/ssl/calibre.
Learning representations for multimedia content is critical for multimedia recommendation. Current representation learning methods roughly fall into two groups: (1) using the historical interactions to create ID embeddings of users and items, and (2) treating multi-modal data as the side information of items to enrich their ID embeddings. Each user-item interaction offers the supervisory signal to optimize the representation learning by the traditional supervised learning paradigm. Due to the overlook of the multi-modal patterns ( e.g. , co-occurrence of visual, acoustic, textual features in micro-videos a user saw before, and her behavioral features) hidden in the data, these methods are insufficient to create powerful representations and obtain satisfactory recommendation accuracy. To capture multi-modal patterns in the data itself, we go beyond the supervised learning paradigm, and incorporate the idea of self-supervised learning (SSL) into multimedia recommendation. Specifically, SSL consists of two components: (1) data augmentation upon multi-modal contents, where we design three operators — feature dropout (FD), feature masking (FM), feature fine and coarse spaces (FAC) — to generate multiple views of individual items; and (2) contrastive learning, which differentiates the views of an item from the others’ to distill additional supervisory signals. Clearly, SSL enables us to explore and exhibit the underlying relations among modalities, thereby resulting in powerful representations. We denote the generic framework by Self-supervised Learning-guided Multimedia Recommendation (SLMRec). Extensive experiments are performed on three real-world datasets, showing that SLMRec achieves significant improvements over several state-of-the-art baselines like LightGCN [1], MMGCN [2]. Further analysis shows how SSL affects recommendation performance.
Graph Self-Supervised Learning (SSL) has emerged as a pivotal area of research in recent years. By engaging in pretext tasks to learn the intricate topological structures and properties of graphs using unlabeled data, these graph SSL models achieve enhanced performance, improved generalization, and heightened robustness. Despite the remarkable achievements of these graph SSL methods, their current implementation poses significant challenges for beginners and practitioners due to the complex nature of graph structures, inconsistent evaluation metrics, and concerns regarding reproducibility hinder further progress in this field. Recognizing the growing interest within the research community, there is an urgent need for a comprehensive, beginner-friendly, and accessible toolkit consisting of the most representative graph SSL algorithms. To address these challenges, we present a Graph SSL toolkit named PyG-SSL, which is built upon PyTorch and is compatible with various deep learning and scientific computing backends. Within the toolkit, we offer a unified framework encompassing dataset loading, hyper-parameter configuration, model training, and comprehensive performance evaluation for diverse downstream tasks. Moreover, we provide beginner-friendly tutorials and the best hyper-parameters of each graph SSL algorithm on different graph datasets, facilitating the reproduction of results. The GitHub repository of the library is https://github.com/iDEA-iSAIL-Lab-UIUC/pyg-ssl.
Understanding the traffic dynamics in networks is a core capability for automated systems to monitor and analyze networking behaviors, reducing expensive human efforts and economic risks through tasks such as traffic classification, congestion prediction, and attack detection. However, it is still challenging to accurately model network traffic with machine learning approaches in an efficient and broadly applicable manner. Task-specific models trained from scratch are used for different networking applications, which limits the efficiency of model development and generalization of model deployment. Furthermore, while networking data is abundant, high-quality task-specific labels are often insufficient for training individual models. Large-scale self-supervised learning on unlabeled data provides a natural pathway for tackling these challenges. We propose to pre-train a general-purpose machine learning model to capture traffic dynamics with only traffic data from NetFlow records, with the goal of fine-tuning for different downstream tasks with small amount of labels. Our presented NetFlowGen framework goes beyond a proof-of-concept for network traffic pre-training and addresses specific challenges such as unifying network feature representations, learning from large unlabeled traffic data volume, and testing on real downstream tasks in DDoS attack detection. Experiments demonstrate promising results of our pre-training framework on capturing traffic dynamics and adapting to different networking tasks.
This study proposes a knowledge distillation algorithm based on large language models and feature alignment, aiming to effectively transfer the knowledge of large pre-trained models into lightweight student models, thereby reducing computational costs while maintaining high model performance. Different from the traditional soft label distillation method, this method introduces a multi-layer feature alignment strategy to deeply align the intermediate features and attention mechanisms of the teacher model and the student model, maximally retaining the semantic expression ability and context modeling ability of the teacher model. In terms of method design, a multi-task loss function is constructed, including feature matching loss, attention alignment loss, and output distribution matching loss, to ensure multi-level information transfer through joint optimization. The experiments were comprehensively evaluated on the GLUE data set and various natural language processing tasks. The results show that the proposed model performs very close to the state-of-the-art GPT-4 model in terms of evaluation indicators such as perplexity, BLEU, ROUGE, and CER. At the same time, it far exceeds baseline models such as DeBERTa, XLNet, and GPT-3, showing significant performance improvements and computing efficiency advantages. Research results show that the feature alignment distillation strategy is an effective model compression method that can significantly reduce computational overhead and storage requirements while maintaining model capabilities. Future research can be further expanded in the directions of self-supervised learning, cross-modal feature alignment, and multi-task transfer learning to provide more flexible and efficient solutions for the deployment and optimization of deep learning models.
Spherical Sliced-Wasserstein (SSW) has recently been proposed to measure the discrepancy between spherical data distributions in various fields, such as geology, medical domains, computer vision, and deep representation learning. However, in the original SSW, all projection directions are treated equally, which is too idealistic and cannot accurately reflect the importance of different projection directions for various data distributions. To address this issue, we propose a novel data-adaptive Discriminative Spherical Sliced-Wasserstein (DSSW) distance, which utilizes a projected energy function to determine the discriminative projection direction for SSW. In our new DSSW, we introduce two types of projected energy functions to generate the weights for projection directions with complete theoretical guarantees. The first type employs a non-parametric deterministic function that transforms the projected Wasserstein distance into its corresponding weight in each projection direction. This improves the performance of the original SSW distance with negligible additional computational overhead. The second type utilizes a neural network-induced function that learns the projection direction weight through a parameterized neural network based on data projections. This further enhances the performance of the original SSW distance with less extra computational overhead. Finally, we evaluate the performance of our proposed DSSW by comparing it with several state-of-the-art methods across a variety of machine learning tasks, including gradient flows, density estimation on real earth data, and self-supervised learning.
In recent years, researchers have attempted to exploit social relations to improve the performance in recommendation systems. Generally, most existing social recommendation methods heavily depends on substantial domain knowledge and expertise in primary recommendation tasks for designing useful auxiliary tasks. Meanwhile, Self-Supervised Learning (SSL) recently has received considerable attention in the field of recommendation, since it can provide self-supervision signals in assisting the improvement of target recommendation systems by constructing self-supervised auxiliary tasks from raw data without human-annotated labels. Despite the great success, these SSL-based social recommendations are insufficient to adaptively balance various self-supervised auxiliary tasks, since assigning equal weights on various auxiliary tasks can result in sub-optimal recommendation performance, where different self-supervised auxiliary tasks may contribute differently to improving the primary social recommendation across different datasets. To address this issue, in this work, we propose Adaptive Self-supervised Learning for Social Recommendations (AdasRec) by taking advantage of various self-supervised auxiliary tasks. More specifically, an adaptive weighting mechanism is proposed to learn adaptive weights for various self-supervised auxiliary tasks, so as to balance the contribution of such self-supervised auxiliary tasks for enhancing representation learning in social recommendations. The adaptive weighting mechanism is used to assign different weights on auxiliary tasks to achieve an overall weighting of the entire auxiliary tasks and ultimately assist the primary recommendation task, achieved by a meta learning optimization problem with an adaptive weighting network. Comprehensive experiments on various real-world datasets are constructed to verify the effectiveness of our proposed method.
Masked point modeling methods have recently achieved great success in self-supervised learning for point cloud data. However, these methods are sensitive to rotations and often exhibit sharp performance drops when encountering rotational variations. In this paper, we propose a novel Rotation-Invariant Masked AutoEncoders (RI-MAE) to address two major challenges: 1) achieving rotation-invariant latent representations, and 2) facilitating self-supervised reconstruction in a rotation-invariant manner. For the first challenge, we introduce RI-Transformer, which features disentangled geometry content, rotation-invariant relative orientation and position embedding mechanisms for constructing rotation-invariant point cloud latent space. For the second challenge, a novel dual-branch student-teacher architecture is devised. It enables the self-supervised learning via the reconstruction of masked patches within the learned rotation-invariant latent space. Each branch is based on an RI-Transformer, and they are connected with an additional RI-Transformer predictor. The teacher encodes all point patches, while the student solely encodes unmasked ones. Finally, the predictor predicts the latent features of the masked patches using the output latent embeddings from the student, supervised by the outputs from the teacher. Extensive experiments demonstrate that our method is robust to rotations, achieving the state-of-the-art performance on various downstream tasks. Our code is available at https://github.com/kunmingsu07/RI-MAE.
AI-synthesized voice technology has the potential to create realistic human voices for beneficial applications, but it can also be misused for malicious purposes. While existing AI-synthesized voice detection models excel in intra-domain evaluation, they face challenges in generalizing across different domains, potentially becoming obsolete as new voice generators emerge. Current solutions use diverse data and advanced machine learning techniques (e.g., domain-invariant representation, self-supervised learning), but are limited by predefined vocoders and sensitivity to factors like background noise and speaker identity. In this work, we introduce an innovative disentanglement framework aimed at extracting domain-agnostic artifact features related to vocoders. Utilizing these features, we enhance model learning in a flat loss landscape, enabling escape from suboptimal solutions and improving generalization. Extensive experiments on benchmarks show our approach outperforms state-of-the-art methods, achieving up to 5.12% improvement in the equal error rate metric in intra-domain and 7.59% in cross-domain evaluations.