Foundation deep learning (DL) models are general models, designed to learn general, robust and adaptable representations of their target modality, enabling finetuning across a range of downstream tasks. These models are pretrained on large, unlabeled datasets using self-supervised learning (SSL). Foundation models have demonstrated better generalization than traditional supervised approaches, a critical requirement for wireless communications where the dynamic environment demands model adaptability. In this work, we propose and demonstrate the effectiveness of a Vision Transformer (ViT) as a radio foundation model for spectrogram learning. We introduce a Masked Spectrogram Modeling (MSM) approach to pretrain the ViT in a self-supervised fashion. We evaluate the ViT-based foundation model on two downstream tasks: Channel State Information (CSI)-based Human Activity sensing and Spectrogram Segmentation. Experimental results demonstrate competitive performance to supervised training while generalizing across diverse domains. Notably, the pretrained ViT model outperforms a four-times larger model that is trained from scratch on the spectrogram segmentation task, while requiring significantly less training time, and achieves competitive performance on the CSI-based human activity sensing task. This work demonstrates the effectiveness of ViT with MSM for pretraining as a promising technique for scalable foundation model development in future 6G networks.
Foundational deep learning (DL) models are general models, trained on large, diverse, and unlabelled datasets, typically using self-supervised learning techniques have led to significant advancements especially in natural language processing. These pretrained models can be fine-tuned for related downstream tasks, offering faster development and reduced training costs, while often achieving improved performance. In this work, we introduce Masked Spectrogram Modeling, a novel self-supervised learning approach for pretraining foundational DL models on radio signals. Adopting a Convolutional LSTM architecture for efficient spatio-temporal processing, we pretrain the model with an unlabelled radio dataset collected from over-the-air measurements. Subsequently, the pretrained model is fine-tuned for two downstream tasks: spectrum forecasting and segmentation. Experimental results demonstrate that our methodology achieves competitive performance in both forecasting accuracy and segmentation, validating its effectiveness for developing foundational radio models.
Self-supervised learning (SSL) is a data-driven learning approach that utilizes the innate structure of the data to guide the learning process. In contrast to supervised learning, which depends on external labels, SSL utilizes the inherent characteristics of the data to produce its own supervisory signal. However, one frequent issue with SSL methods is representation collapse, where the model outputs a constant input-invariant feature representation. This issue hinders the potential application of SSL methods to new data modalities, as trying to avoid representation collapse wastes researchers' time and effort. This paper introduces a novel SSL algorithm for time-series data called Prediction of Functionals from Masked Latents (PFML). Instead of predicting masked input signals or their latent representations directly, PFML operates by predicting statistical functionals of the input signal corresponding to masked embeddings, given a sequence of unmasked embeddings. The algorithm is designed to avoid representation collapse, rendering it straightforwardly applicable to different time-series data domains, such as novel sensor modalities in clinical data. We demonstrate the effectiveness of PFML through complex, real-life classification tasks across three different data modalities: infant posture and movement classification from multi-sensor inertial measurement unit data, emotion recognition from speech data, and sleep stage classification from EEG data. The results show that PFML is superior to a conceptually similar SSL method and a contrastive learning-based SSL method. Additionally, PFML is on par with the current state-of-the-art SSL method, while also being conceptually simpler and without suffering from representation collapse.
Egocentric videos provide valuable insights into human interactions with the physical world, which has sparked growing interest in the computer vision and robotics communities. A critical challenge in fully understanding the geometry and dynamics of egocentric videos is dense scene reconstruction. However, the lack of high-quality labeled datasets in this field has hindered the effectiveness of current supervised learning methods. In this work, we aim to address this issue by exploring an self-supervised dynamic scene reconstruction approach. We introduce EgoMono4D, a novel model that unifies the estimation of multiple variables necessary for Egocentric Monocular 4D reconstruction, including camera intrinsic, camera poses, and video depth, all within a fast feed-forward framework. Starting from pretrained single-frame depth and intrinsic estimation model, we extend it with camera poses estimation and align multi-frame results on large-scale unlabeled egocentric videos. We evaluate EgoMono4D in both in-domain and zero-shot generalization settings, achieving superior performance in dense pointclouds sequence reconstruction compared to all baselines. EgoMono4D represents the first attempt to apply self-supervised learning for pointclouds sequence reconstruction to the label-scarce egocentric field, enabling fast, dense, and generalizable reconstruction. The interactable visualization, code and trained models are released https://egomono4d.github.io/
Accurate left atrium (LA) segmentation from pre-operative scans is crucial for diagnosing atrial fibrillation, treatment planning, and supporting surgical interventions. While deep learning models are key in medical image segmentation, they often require extensive manually annotated data. Foundation models trained on larger datasets have reduced this dependency, enhancing generalizability and robustness through transfer learning. We explore DINOv2, a self-supervised learning vision transformer trained on natural images, for LA segmentation using MRI. The challenges for LA's complex anatomy, thin boundaries, and limited annotated data make accurate segmentation difficult before & during the image-guided intervention. We demonstrate DINOv2's ability to provide accurate & consistent segmentation, achieving a mean Dice score of .871 & a Jaccard Index of .792 for end-to-end fine-tuning. Through few-shot learning across various data sizes & patient counts, DINOv2 consistently outperforms baseline models. These results suggest that DINOv2 effectively adapts to MRI with limited data, highlighting its potential as a competitive tool for segmentation & encouraging broader use in medical imaging.
In view of the fact that semi- and self-supervised learning share a fundamental principle, effectively modeling knowledge from unlabeled data, various semi-supervised semantic segmentation methods have integrated representative self-supervised learning paradigms for further regularization. However, the potential of the state-of-the-art generative self-supervised paradigm, masked image modeling, has been scarcely studied. This paradigm learns the knowledge through establishing connections between the masked and visible parts of masked image, during the pixel reconstruction process. By inheriting and extending this insight, we successfully leverage masked image modeling to boost semi-supervised semantic segmentation. Specifically, we introduce a novel class-wise masked image modeling that independently reconstructs different image regions according to their respective classes. In this way, the mask-induced connections are established within each class, mitigating the semantic confusion that arises from plainly reconstructing images in basic masked image modeling. To strengthen these intra-class connections, we further develop a feature aggregation strategy that minimizes the distances between features corresponding to the masked and visible parts within the same class. Additionally, in semantic space, we explore the application of masked image modeling to enhance regularization. Extensive experiments conducted on well-known benchmarks demonstrate that our approach achieves state-of-the-art performance. The code will be available at https://github.com/haoxt/S4MIM.
This study introduces a novel self-supervised learning approach for volumetric segmentation of defect indications captured by phased array ultrasonic testing data from Carbon Fiber Reinforced Polymers (CFRPs). By employing this self-supervised method, defect segmentation is achieved automatically without the need for labelled training data or examples of defects. The approach has been tested using artificially induced defects, including back-drilled holes and Polytetrafluoroethylene (PTFE) inserts, to mimic different defect responses. Additionally, it has been evaluated on stepped geometries with varying thickness, demonstrating impressive generalization across various test scenarios. Minimal preprocessing requirements are needed, with no removal of geometric features or Time-Compensated Gain (TCG) necessary for applying the methodology. The model's performance was evaluated for defect detection, in-plane and through-thickness localisation, as well as defect sizing. All defects were consistently detected with thresholding and different processing steps able to remove false positive indications for a 100% detection accuracy. Defect sizing aligns with the industrial standard 6 dB amplitude drop method, with a Mean Absolute Error (MAE) of 1.41 mm. In-plane and through-thickness localisation yielded comparable results, with MAEs of 0.37 and 0.26 mm, respectively. Visualisations are provided to illustrate how this approach can be utilised to generate digital twins of components.
Coronary artery disease (CAD), one of the leading causes of mortality worldwide, necessitates effective risk assessment strategies, with coronary artery calcium (CAC) scoring via computed tomography (CT) being a key method for prevention. Traditional methods, primarily based on UNET architectures implemented on pre-built models, face challenges like the scarcity of annotated CT scans containing CAC and imbalanced datasets, leading to reduced performance in segmentation and scoring tasks. In this study, we address these limitations by incorporating the self-supervised learning (SSL) technique of DINO (self-distillation with no labels), which trains without requiring CAC-specific annotations, enhancing its robustness in generating distinct features. The DINO-LG model, which leverages label guidance to focus on calcified areas, achieves significant improvements, with a sensitivity of 89% and specificity of 90% for detecting CAC-containing CT slices, compared to the standard DINO model's sensitivity of 79% and specificity of 77%. Additionally, false-negative and false-positive rates are reduced by 49% and 59%, respectively, instilling greater confidence in clinicians when ruling out calcification in low-risk patients and minimizing unnecessary imaging reviews by radiologists. Further, CAC scoring and segmentation tasks are conducted using a basic UNET architecture, applied specifically to CT slices identified by the DINO-LG model as containing calcified areas. This targeted approach enhances CAC scoring accuracy by feeding the UNET model with relevant slices, significantly improving diagnostic precision, reducing both false positives and false negatives, and ultimately lowering overall healthcare costs by minimizing unnecessary tests and treatments, presenting a valuable advancement in CAD risk assessment.
Biophysical model fitting plays a key role in obtaining quantitative parameters from physiological signals and images. However, the model complexity for molecular magnetic resonance imaging (MRI) often translates into excessive computation time, which makes clinical use impractical. Here, we present a generic computational approach for solving the parameter extraction inverse problem posed by ordinary differential equation (ODE) modeling coupled with experimental measurement of the system dynamics. This is achieved by formulating a numerical ODE solver to function as a step-wise analytical one, thereby making it compatible with automatic differentiation-based optimization. This enables efficient gradient-based model fitting, and provides a new approach to parameter quantification based on self-supervised learning from a single data observation. The neural-network-based train-by-fit pipeline was used to quantify semisolid magnetization transfer (MT) and chemical exchange saturation transfer (CEST) amide proton exchange parameters in the human brain, in an in-vivo molecular MRI study (n = 4). The entire pipeline of the first whole brain quantification was completed in 18.3 $\pm$ 8.3 minutes. Reusing the single-subject-trained network for inference in new subjects took 1.0 $\pm$ 0.2 s, to provide results in agreement with literature values and scan-specific fit results.
Recent self-supervised learning (SSL) methods have demonstrated impressive results in learning visual representations from unlabeled remote sensing images. However, most remote sensing images predominantly consist of scenographic scenes containing multiple ground objects without explicit foreground targets, which limits the performance of existing SSL methods that focus on foreground targets. This raises the question: Is there a method that can automatically aggregate similar objects within scenographic remote sensing images, thereby enabling models to differentiate knowledge embedded in various geospatial patterns for improved feature representation? In this work, we present the Pattern Integration and Enhancement Vision Transformer (PIEViT), a novel self-supervised learning framework designed specifically for remote sensing imagery. PIEViT utilizes a teacher-student architecture to address both image-level and patch-level tasks. It employs the Geospatial Pattern Cohesion (GPC) module to explore the natural clustering of patches, enhancing the differentiation of individual features. The Feature Integration Projection (FIP) module further refines masked token reconstruction using geospatially clustered patches. We validated PIEViT across multiple downstream tasks, including object detection, semantic segmentation, and change detection. Experiments demonstrated that PIEViT enhances the representation of internal patch features, providing significant improvements over existing self-supervised baselines. It achieves excellent results in object detection, land cover classification, and change detection, underscoring its robustness, generalization, and transferability for remote sensing image interpretation tasks.
Colorectal cancer (CRC) is one of the few cancers that have an established dysplasia-carcinoma sequence that benefits from screening. Everyone over 50 years of age in Canada is eligible for CRC screening. About 20\% of those people will undergo a biopsy for a pre-neoplastic polyp and, in many cases, multiple polyps. As such, these polyp biopsies make up the bulk of a pathologist's workload. Developing an efficient computational model to help screen these polyp biopsies can improve the pathologist's workflow and help guide their attention to critical areas on the slide. DL models face significant challenges in computational pathology (CPath) because of the gigapixel image size of whole-slide images and the scarcity of detailed annotated datasets. It is, therefore, crucial to leverage self-supervised learning (SSL) methods to alleviate the burden and cost of data annotation. However, current research lacks methods to apply SSL frameworks to analyze pathology data effectively. This paper aims to propose an optimized Barlow Twins framework for colorectal polyps screening. We adapt its hyperparameters, augmentation strategy and encoder to the specificity of the pathology data to enhance performance. Additionally, we investigate the best Field of View (FoV) for colorectal polyps screening and propose a new benchmark dataset for CRC screening, made of four types of colorectal polyps and normal tissue, by performing downstream tasking on MHIST and NCT-CRC-7K datasets. Furthermore, we show that the SSL representations are more meaningful and qualitative than the supervised ones and that Barlow Twins benefits from the Swin Transformer when applied to pathology data. Codes are avaialble from https://github.com/AtlasAnalyticsLab/PathBT.
Accurate prediction of cardiovascular diseases remains imperative for early diagnosis and intervention, necessitating robust and precise predictive models. Recently, there has been a growing interest in multi-modal learning for uncovering novel insights not available through uni-modal datasets alone. By combining cardiac magnetic resonance images, electrocardiogram signals, and available medical information, our approach enables the capture of holistic status about individuals' cardiovascular health by leveraging shared information across modalities. Integrating information from multiple modalities and benefiting from self-supervised learning techniques, our model provides a comprehensive framework for enhancing cardiovascular disease prediction with limited annotated datasets. We employ a masked autoencoder to pre-train the electrocardiogram ECG encoder, enabling it to extract relevant features from raw electrocardiogram data, and an image encoder to extract relevant features from cardiac magnetic resonance images. Subsequently, we utilize a multi-modal contrastive learning objective to transfer knowledge from expensive and complex modality, cardiac magnetic resonance image, to cheap and simple modalities such as electrocardiograms and medical information. Finally, we fine-tuned the pre-trained encoders on specific predictive tasks, such as myocardial infarction. Our proposed method enhanced the image information by leveraging different available modalities and outperformed the supervised approach by 7.6% in balanced accuracy.
Ultrasound (US) image stitching can expand the field-of-view (FOV) by combining multiple US images from varied probe positions. However, registering US images with only partially overlapping anatomical contents is a challenging task. In this work, we introduce SynStitch, a self-supervised framework designed for 2DUS stitching. SynStitch consists of a synthetic stitching pair generation module (SSPGM) and an image stitching module (ISM). SSPGM utilizes a patch-conditioned ControlNet to generate realistic 2DUS stitching pairs with known affine matrix from a single input image. ISM then utilizes this synthetic paired data to learn 2DUS stitching in a supervised manner. Our framework was evaluated against multiple leading methods on a kidney ultrasound dataset, demonstrating superior 2DUS stitching performance through both qualitative and quantitative analyses. The code will be made public upon acceptance of the paper.
Large language models (LMs) are typically adapted to improve performance on new contexts (\eg text prompts that define new tasks or domains) through fine-tuning or prompting. However, there is an accuracy compute tradeoff -- fine-tuning incurs significant training cost and prompting increases inference overhead. We introduce $GenerativeAdapter$, an effective and efficient adaptation method that directly maps new contexts to low-rank LM adapters, thereby significantly reducing inference overhead with no need for finetuning. The adapter generator is trained via self-supervised learning, and can be used to adapt a single frozen LM for any new task simply by mapping the associated task or domain context to a new adapter. We apply $GenerativeAdapter$ to two pretrained LMs (Mistral-7B-Instruct and Llama2-7B-Chat) and evaluate the adapted models in three adaption scenarios: knowledge acquisition from documents, learning from demonstrations, and personalization for users. In StreamingQA, our approach is effective in injecting knowledge into the LM's parameters, achieving a 63.5% improvement in F1 score over the model with supervised fine-tuning (from $19.5$ to $31.5$) for contexts as long as 32K tokens. In the MetaICL in-context learning evaluation, our method achieves an average accuracy of $44.9$ across 26 tasks, outperforming the base model. On MSC, our method proves to be highly competitive in memorizing user information from conversations with a 4x reduction in computation and memory costs compared to prompting with full conversation history. Together, these results suggest that $GenerativeAdapter$ should allow for general adaption to a wide range of different contexts.
Contrastive learning has been a leading paradigm for self-supervised learning, but it is widely observed that it comes at the price of sacrificing useful features (\eg colors) by being invariant to data augmentations. Given this limitation, there has been a surge of interest in equivariant self-supervised learning (E-SSL) that learns features to be augmentation-aware. However, even for the simplest rotation prediction method, there is a lack of rigorous understanding of why, when, and how E-SSL learns useful features for downstream tasks. To bridge this gap between practice and theory, we establish an information-theoretic perspective to understand the generalization ability of E-SSL. In particular, we identify a critical explaining-away effect in E-SSL that creates a synergy between the equivariant and classification tasks. This synergy effect encourages models to extract class-relevant features to improve its equivariant prediction, which, in turn, benefits downstream tasks requiring semantic features. Based on this perspective, we theoretically analyze the influence of data transformations and reveal several principles for practical designs of E-SSL. Our theory not only aligns well with existing E-SSL methods but also sheds light on new directions by exploring the benefits of model equivariance. We believe that a theoretically grounded understanding on the role of equivariance would inspire more principled and advanced designs in this field. Code is available at https://github.com/kaotty/Understanding-ESSL.
Graph representation learning (GRL) is a fundamental task in machine learning, aiming to encode high-dimensional graph-structured data into low-dimensional vectors. Self-supervised learning (SSL) methods are widely used in GRL because they can avoid expensive human annotation. In this work, we propose a novel Subgraph Gaussian Embedding Contrast (SGEC) method. Our approach introduces a subgraph Gaussian embedding module, which adaptively maps subgraphs to a structured Gaussian space, ensuring the preservation of graph characteristics while controlling the distribution of generated subgraphs. We employ optimal transport distances, including Wasserstein and Gromov-Wasserstein distances, to effectively measure the similarity between subgraphs, enhancing the robustness of the contrastive learning process. Extensive experiments across multiple benchmarks demonstrate that SGEC outperforms or presents competitive performance against state-of-the-art approaches. Our findings provide insights into the design of SSL methods for GRL, emphasizing the importance of the distribution of the generated contrastive pairs.
Early identification of stroke is crucial for intervention, requiring reliable models. We proposed an efficient retinal image representation together with clinical information to capture a comprehensive overview of cardiovascular health, leveraging large multimodal datasets for new medical insights. Our approach is one of the first contrastive frameworks that integrates graph and tabular data, using vessel graphs derived from retinal images for efficient representation. This method, combined with multimodal contrastive learning, significantly enhances stroke prediction accuracy by integrating data from multiple sources and using contrastive learning for transfer learning. The self-supervised learning techniques employed allow the model to learn effectively from unlabeled data, reducing the dependency on large annotated datasets. Our framework showed an AUROC improvement of 3.78% from supervised to self-supervised approaches. Additionally, the graph-level representation approach achieved superior performance to image encoders while significantly reducing pre-training and fine-tuning runtimes. These findings indicate that retinal images are a cost-effective method for improving cardiovascular disease predictions and pave the way for future research into retinal and cerebral vessel connections and the use of graph-based retinal vessel representations.
The growing demand for accurate and equitable AI models in digital dermatology faces a significant challenge: the lack of diverse, high-quality labeled data. In this work, we investigate the potential of domain-specific foundation models for dermatology in addressing this challenge. We utilize self-supervised learning (SSL) techniques to pre-train models on a dataset of over 240,000 dermatological images from public and private collections. Our study considers several SSL methods and compares the resulting foundation models against domain-agnostic models like those pre-trained on ImageNet and state-of-the-art models such as MONET across 12 downstream tasks. Unlike previous research, we emphasize the development of smaller models that are more suitable for resource-limited clinical settings, facilitating easier adaptation to a broad range of use cases. Results show that models pre-trained in this work not only outperform general-purpose models but also approach the performance of models 50 times larger on clinically relevant diagnostic tasks. To promote further research in this direction, we publicly release both the training code and the foundation models, which can benefit clinicians in dermatological applications.
Self-supervised learning (SSL) has grown in interest within the speech processing community, since it produces representations that are useful for many downstream tasks. SSL uses global and contextual methods to produce robust representations, where SSL even outperforms supervised models. Most self-supervised approaches, however, are limited to embedding information about, i.e., the phonemes, speaker identity, and emotion, into the extracted representations, where they become invariant to background sounds due to contrastive and auto-regressive learning. This is limiting because many downstream tasks leverage noise information to function accurately. Therefore, we propose a pre-training framework that learns information pertaining to background noise in a supervised manner, while jointly embedding speech information using a self-supervised strategy. We experiment with multiple encoders and show that our framework is useful for perceptual speech quality estimation, which relies on background cues. Our results show that the proposed approach improves performance with fewer parameters, in comparison to multiple baselines.
In this paper, we propose a novel Self-Supervised-Learning scheme to train rhythm analysis systems and instantiate it for few-shot beat tracking. Taking inspiration from the Contrastive Predictive Coding paradigm, we propose to train a Log-Mel-Spectrogram Transformer encoder to contrast observations at times separated by hypothesized beat intervals from those that are not. We do this without the knowledge of ground-truth tempo or beat positions, as we rely on the local maxima of a Predominant Local Pulse function, considered as a proxy for Tatum positions, to define candidate anchors, candidate positives (located at a distance of a power of two from the anchor) and negatives (remaining time positions). We show that a model pre-trained using this approach on the unlabeled FMA, MTT and MTG-Jamendo datasets can successfully be fine-tuned in the few-shot regime, i.e. with just a few annotated examples to get a competitive beat-tracking performance.