Normalization layers are ubiquitous in modern neural networks and have long been considered essential. This work demonstrates that Transformers without normalization can achieve the same or better performance using a remarkably simple technique. We introduce Dynamic Tanh (DyT), an element-wise operation $DyT($x$) = \tanh(\alpha $x$)$, as a drop-in replacement for normalization layers in Transformers. DyT is inspired by the observation that layer normalization in Transformers often produces tanh-like, $S$-shaped input-output mappings. By incorporating DyT, Transformers without normalization can match or exceed the performance of their normalized counterparts, mostly without hyperparameter tuning. We validate the effectiveness of Transformers with DyT across diverse settings, ranging from recognition to generation, supervised to self-supervised learning, and computer vision to language models. These findings challenge the conventional understanding that normalization layers are indispensable in modern neural networks, and offer new insights into their role in deep networks.
Recent advances in self-supervised learning for Vision Transformers (ViTs) have fueled breakthroughs in remote sensing (RS) foundation models. However, the quadratic complexity of self-attention poses a significant barrier to scalability, particularly for large models and high-resolution images. While the linear-complexity Mamba architecture offers a promising alternative, existing RS applications of Mamba remain limited to supervised tasks on small, domain-specific datasets. To address these challenges, we propose RoMA, a framework that enables scalable self-supervised pretraining of Mamba-based RS foundation models using large-scale, diverse, unlabeled data. RoMA enhances scalability for high-resolution images through a tailored auto-regressive learning strategy, incorporating two key innovations: 1) a rotation-aware pretraining mechanism combining adaptive cropping with angular embeddings to handle sparsely distributed objects with arbitrary orientations, and 2) multi-scale token prediction objectives that address the extreme variations in object scales inherent to RS imagery. Systematic empirical studies validate that Mamba adheres to RS data and parameter scaling laws, with performance scaling reliably as model and data size increase. Furthermore, experiments across scene classification, object detection, and semantic segmentation tasks demonstrate that RoMA-pretrained Mamba models consistently outperform ViT-based counterparts in both accuracy and computational efficiency. The source code and pretrained models will be released at https://github.com/MiliLab/RoMA.
Electroencephalography (EEG) serves as an essential diagnostic tool in neurology; however, its accurate manual interpretation is a time-intensive process that demands highly specialized expertise, which remains relatively scarce and not consistently accessible. To address these limitations, the implementation of automated pre-screening and analysis systems for EEG data holds considerable promise. Advances in self-supervised learning made it possible to pre-train complex deep learning architectures on large volumes of unlabeled EEG data to learn generalizable representations, that can later be used to enhance performance on multiple tasks while needing less downstream data. In the present paper, we introduce BioSerenity-E1, the first of a family of self-supervised foundation models for clinical EEG applications that combines spectral tokenization with masked prediction to achieve state-of-the-art performance across relevant diagnostic tasks. The two-phase self-supervised pretraining framework initially acquires compressed EEG representations via a transformer-based VQ-VAE architecture designed to reconstruct log-multitaper spectral projections, then implements extensive (70% block) masked token prediction to force the model to learn complex spatiotemporal dependencies in EEG signals. BioSerenity-E1 achieves strong performance across three clinical tasks, either in line or above state-of-the-art methods: seizure detection (AUROC = 0.926, Sensitivity = 0.909), normal/abnormal classification (AUPRC = 0.970 on proprietary data; 0.910 on TUH-Abnormal), and multiclass pathology differentiation on unbalanced data (Weighted F1 = 0.730). The utility of BioSerenity-E1 is further confirmed in low-data regimes scenarios, showing clear improvements in AUPRC (from +2% to 17%) when trained on less than 10% of the available data.
This work aims to tackle the Parkinson's disease (PD) detection problem from the speech signal in a bilingual setting by proposing an ad-hoc dual-head deep neural architecture for type-based binary classification. One head is specialized for diadochokinetic patterns. The other head looks for natural speech patterns present in continuous spoken utterances. Only one of the two heads is operative accordingly to the nature of the input. Speech representations are extracted from self-supervised learning (SSL) models and wavelet transforms. Adaptive layers, convolutional bottlenecks, and contrastive learning are exploited to reduce variations across languages. Our solution is assessed against two distinct datasets, EWA-DB, and PC-GITA, which cover Slovak and Spanish languages, respectively. Results indicate that conventional models trained on a single language dataset struggle with cross-linguistic generalization, and naive combinations of datasets are suboptimal. In contrast, our model improves generalization on both languages, simultaneously.
Although Vision Transformers (ViTs) have recently demonstrated superior performance in medical imaging problems, they face explainability issues similar to previous architectures such as convolutional neural networks. Recent research efforts suggest that attention maps, which are part of decision-making process of ViTs can potentially address the explainability issue by identifying regions influencing predictions, especially in models pretrained with self-supervised learning. In this work, we compare the visual explanations of attention maps to other commonly used methods for medical imaging problems. To do so, we employ four distinct medical imaging datasets that involve the identification of (1) colonic polyps, (2) breast tumors, (3) esophageal inflammation, and (4) bone fractures and hardware implants. Through large-scale experiments on the aforementioned datasets using various supervised and self-supervised pretrained ViTs, we find that although attention maps show promise under certain conditions and generally surpass GradCAM in explainability, they are outperformed by transformer-specific interpretability methods. Our findings indicate that the efficacy of attention maps as a method of interpretability is context-dependent and may be limited as they do not consistently provide the comprehensive insights required for robust medical decision-making.
Classical shape descriptors such as Heat Kernel Signature (HKS), Wave Kernel Signature (WKS), and Signature of Histograms of OrienTations (SHOT), while widely used in shape analysis, exhibit sensitivity to mesh connectivity, sampling patterns, and topological noise. While differential geometry offers a promising alternative through its theory of differential invariants, which are theoretically guaranteed to be robust shape descriptors, the computation of these invariants on discrete meshes often leads to unstable numerical approximations, limiting their practical utility. We present a self-supervised learning approach for extracting geometric features from 3D surfaces. Our method combines synthetic data generation with a neural architecture designed to learn sampling-invariant features. By integrating our features into existing shape correspondence frameworks, we demonstrate improved performance on standard benchmarks including FAUST, SCAPE, TOPKIDS, and SHREC'16, showing particular robustness to topological noise and partial shapes.
In this work we present a novel approach for unsupervised multi-graph matching, which applies to problems for which a Gaussian distribution of keypoint features can be assumed. We leverage cycle consistency as loss for self-supervised learning, and determine Gaussian parameters through Bayesian Optimization, yielding a highly efficient approach that scales to large datasets. Our fully unsupervised approach enables us to reach the accuracy of state-of-the-art supervised methodology for the use case of annotating cell nuclei in 3D microscopy images of the worm C. elegans. To this end, our approach yields the first unsupervised atlas of C. elegans, i.e. a model of the joint distribution of all of its cell nuclei, without the need for any ground truth cell annotation. This advancement enables highly efficient annotation of cell nuclei in large microscopy datasets of C. elegans. Beyond C. elegans, our approach offers fully unsupervised construction of cell-level atlases for any model organism with a stereotyped cell lineage, and thus bears the potential to catalyze respective comparative developmental studies in a range of further species.
Accurate depth and camera pose estimation is essential for achieving high-quality 3D visualisations in robotic-assisted surgery. Despite recent advancements in foundation model adaptation to monocular depth estimation of endoscopic scenes via self-supervised learning (SSL), no prior work has explored their use for pose estimation. These methods rely on low rank-based adaptation approaches, which constrain model updates to a low-rank space. We propose Endo-FASt3r, the first monocular SSL depth and pose estimation framework that uses foundation models for both tasks. We extend the Reloc3r relative pose estimation foundation model by designing Reloc3rX, introducing modifications necessary for convergence in SSL. We also present DoMoRA, a novel adaptation technique that enables higher-rank updates and faster convergence. Experiments on the SCARED dataset show that Endo-FASt3r achieves a substantial $10\%$ improvement in pose estimation and a $2\%$ improvement in depth estimation over prior work. Similar performance gains on the Hamlyn and StereoMIS datasets reinforce the generalisability of Endo-FASt3r across different datasets.
Self-supervised learning (SSL) has garnered substantial interest within the machine learning and computer vision communities. Two prominent approaches in SSL include contrastive-based learning and self-distillation utilizing cropping augmentation. Lately, masked image modeling (MIM) has emerged as a more potent SSL technique, employing image inpainting as a pretext task. MIM creates a strong inductive bias toward meaningful spatial and semantic understanding. This has opened up new opportunities for SSL to contribute not only to classification tasks but also to more complex applications like object detection and image segmentation. Building upon this progress, our research paper introduces a scalable and practical SSL approach centered around more challenging pretext tasks that facilitate the acquisition of robust features. Specifically, we leverage multi-scale image reconstruction from randomly masked input images as the foundation for feature learning. Our hypothesis posits that reconstructing high-resolution images enables the model to attend to finer spatial details, particularly beneficial for discerning subtle intricacies within medical images. The proposed SSL features help improve classification performance on the Curated Breast Imaging Subset of Digital Database for Screening Mammography (CBIS-DDSM) dataset. In pathology classification, our method demonstrates a 3\% increase in average precision (AP) and a 1\% increase in the area under the receiver operating characteristic curve (AUC) when compared to state-of-the-art (SOTA) algorithms. Moreover, in mass margins classification, our approach achieves a 4\% increase in AP and a 2\% increase in AUC.
Computational pathology and whole-slide image (WSI) analysis are pivotal in cancer diagnosis and prognosis. However, the ultra-high resolution of WSIs presents significant modeling challenges. Recent advancements in pathology foundation models have improved performance, yet most approaches rely on [CLS] token representation of tile ViT as slide-level inputs (16x16 pixels is refereed as patch and 224x224 pixels as tile). This discards critical spatial details from patch tokens, limiting downstream WSI analysis tasks. We find that leveraging all spatial patch tokens benefits WSI analysis but incurs nearly 200x higher storage and training costs (e.g., 196 tokens in ViT$_{224}$). To address this, we introduce vector quantized (VQ) distillation on patch feature, which efficiently compresses spatial patch tokens using discrete indices and a decoder. Our method reduces token dimensionality from 1024 to 16, achieving a 64x compression rate while preserving reconstruction fidelity. Furthermore, we employ a multi-scale VQ (MSVQ) strategy, which not only enhances VQ reconstruction performance but also serves as a Self-supervised Learning (SSL) supervision for a seamless slide-level pretraining objective. Built upon the quantized patch features and supervision targets of tile via MSVQ, we develop a progressive convolutional module and slide-level SSL to extract representations with rich spatial-information for downstream WSI tasks. Extensive evaluations on multiple datasets demonstrate the effectiveness of our approach, achieving state-of-the-art performance in WSI analysis. Code will be available soon.
Self-supervised learning (SSL) has advanced significantly in visual representation learning, yet comprehensive evaluations of its adversarial robustness remain limited. In this study, we evaluate the adversarial robustness of seven discriminative self-supervised models and one supervised model across diverse tasks, including ImageNet classification, transfer learning, segmentation, and detection. Our findings suggest that discriminative SSL models generally exhibit better robustness to adversarial attacks compared to their supervised counterpart on ImageNet, with this advantage extending to transfer learning when using linear evaluation. However, when fine-tuning is applied, the robustness gap between SSL and supervised models narrows considerably. Similarly, this robustness advantage diminishes in segmentation and detection tasks. We also investigate how various factors might influence adversarial robustness, including architectural choices, training duration, data augmentations, and batch sizes. Our analysis contributes to the ongoing exploration of adversarial robustness in visual self-supervised representation systems.
Multimodal image-tabular learning is gaining attention, yet it faces challenges due to limited labeled data. While earlier work has applied self-supervised learning (SSL) to unlabeled data, its task-agnostic nature often results in learning suboptimal features for downstream tasks. Semi-supervised learning (SemiSL), which combines labeled and unlabeled data, offers a promising solution. However, existing multimodal SemiSL methods typically focus on unimodal or modality-shared features, ignoring valuable task-relevant modality-specific information, leading to a Modality Information Gap. In this paper, we propose STiL, a novel SemiSL tabular-image framework that addresses this gap by comprehensively exploring task-relevant information. STiL features a new disentangled contrastive consistency module to learn cross-modal invariant representations of shared information while retaining modality-specific information via disentanglement. We also propose a novel consensus-guided pseudo-labeling strategy to generate reliable pseudo-labels based on classifier consensus, along with a new prototype-guided label smoothing technique to refine pseudo-label quality with prototype embeddings, thereby enhancing task-relevant information learning in unlabeled data. Experiments on natural and medical image datasets show that STiL outperforms the state-of-the-art supervised/SSL/SemiSL image/multimodal approaches. Our code is publicly available.
This paper focuses on multimodal alignment within the realm of Artificial Intelligence, particularly in text and image modalities. The semantic gap between the textual and visual modality poses a discrepancy problem towards the effectiveness of multi-modalities fusion. Therefore, we introduce Text-Image Joint Embedding Predictive Architecture (TI-JEPA), an innovative pre-training strategy that leverages energy-based model (EBM) framework to capture complex cross-modal relationships. TI-JEPA combines the flexibility of EBM in self-supervised learning to facilitate the compatibility between textual and visual elements. Through extensive experiments across multiple benchmarks, we demonstrate that TI-JEPA achieves state-of-the-art performance on multimodal sentiment analysis task (and potentially on a wide range of multimodal-based tasks, such as Visual Question Answering), outperforming existing pre-training methodologies. Our findings highlight the potential of using energy-based framework in advancing multimodal fusion and suggest significant improvements for downstream applications.
Self-supervised representation learning methods often fail to learn subtle or complex features, which can be dominated by simpler patterns which are much easier to learn. This limitation is particularly problematic in applications to science and engineering, as complex features can be critical for discovery and analysis. To address this, we introduce Split Component Embedding Registration (SpliCER), a novel architecture which splits the image into sections and distils information from each section to guide the model to learn more subtle and complex features without compromising on simpler features. SpliCER is compatible with any self-supervised loss function and can be integrated into existing methods without modification. The primary contributions of this work are as follows: i) we demonstrate that existing self-supervised methods can learn shortcut solutions when simple and complex features are both present; ii) we introduce a novel self-supervised training method, SpliCER, to overcome the limitations of existing methods, and achieve significant downstream performance improvements; iii) we demonstrate the effectiveness of SpliCER in cutting-edge medical and geospatial imaging settings. SpliCER offers a powerful new tool for representation learning, enabling models to uncover complex features which could be overlooked by other methods.
Histopathology image analysis is fundamental to digital pathology, with hematoxylin and eosin (H&E) staining as the gold standard for diagnostic and prognostic assessments. While H&E imaging effectively highlights cellular and tissue structures, it lacks sensitivity to birefringence and tissue anisotropy, which are crucial for assessing collagen organization, fiber alignment, and microstructural alterations--key indicators of tumor progression, fibrosis, and other pathological conditions. To bridge this gap, we propose PolarHE, a dual modality fusion framework that integrates H&E with polarization imaging, leveraging the polarization ability to enhance tissue characterization. Our approach employs a feature decomposition strategy to disentangle common and modality specific features, ensuring effective multimodal representation learning. Through comprehensive validation, our approach significantly outperforms previous methods, achieving an accuracy of 86.70% on the Chaoyang dataset and 89.06% on the MHIST dataset. Moreover, polarization property visualization reveals distinct optical signatures of pathological tissues, highlighting its diagnostic potential. t-SNE visualizations further confirm our model effectively captures both shared and unique modality features, reinforcing the complementary nature of polarization imaging. These results demonstrate that polarization imaging is a powerful and underutilized modality in computational pathology, enriching feature representation and improving diagnostic accuracy. PolarHE establishes a promising direction for multimodal learning, paving the way for more interpretable and generalizable pathology models. Our code will be released after paper acceptance.
Child speech recognition is still an underdeveloped area of research due to the lack of data (especially on non-English languages) and the specific difficulties of this task. Having explored various architectures for child speech recognition in previous work, in this article we tackle recent self-supervised models. We first compare wav2vec 2.0, HuBERT and WavLM models adapted to phoneme recognition in French child speech, and continue our experiments with the best of them, WavLM base+. We then further adapt it by unfreezing its transformer blocks during fine-tuning on child speech, which greatly improves its performance and makes it significantly outperform our base model, a Transformer+CTC. Finally, we study in detail the behaviour of these two models under the real conditions of our application, and show that WavLM base+ is more robust to various reading tasks and noise levels. Index Terms: speech recognition, child speech, self-supervised learning
Moving object segmentation (MOS) on LiDAR point clouds is crucial for autonomous systems like self-driving vehicles. Previous supervised approaches rely heavily on costly manual annotations, while LiDAR sequences naturally capture temporal motion cues that can be leveraged for self-supervised learning. In this paper, we propose \textbf{T}emporal \textbf{O}verlapping \textbf{P}rediction (\textbf{TOP}), a self-supervised pre-training method that alleviate the labeling burden for MOS. \textbf{TOP} explores the temporal overlapping points that commonly observed by current and adjacent scans, and learns spatiotemporal representations by predicting the occupancy states of temporal overlapping points. Moreover, we utilize current occupancy reconstruction as an auxiliary pre-training objective, which enhances the current structural awareness of the model. We conduct extensive experiments and observe that the conventional metric Intersection-over-Union (IoU) shows strong bias to objects with more scanned points, which might neglect small or distant objects. To compensate for this bias, we introduce an additional metric called $\text{mIoU}_{\text{obj}}$ to evaluate object-level performance. Experiments on nuScenes and SemanticKITTI show that \textbf{TOP} outperforms both supervised training-from-scratch baseline and other self-supervised pre-training baselines by up to 28.77\% relative improvement, demonstrating strong transferability across LiDAR setups and generalization to other tasks. Code and pre-trained models will be publicly available upon publication.
Waste classification is crucial for improving processing efficiency and reducing environmental pollution. Supervised deep learning methods are commonly used for automated waste classification, but they rely heavily on large labeled datasets, which are costly and inefficient to obtain. Real-world waste data often exhibit category and style biases, such as variations in camera angles, lighting conditions, and types of waste, which can impact the model's performance and generalization ability. Therefore, constructing a bias-free dataset is essential. Manual labeling is not only costly but also inefficient. While self-supervised learning helps address data scarcity, it still depends on some labeled data and generally results in lower accuracy compared to supervised methods. Unsupervised methods show potential in certain cases but typically do not perform as well as supervised models, highlighting the need for an efficient and cost-effective unsupervised approach. This study presents a novel unsupervised method, Dual-Encoder Contrastive Learning with Multi-Clustering Voting (DECMCV). The approach involves using a pre-trained ConvNeXt model for image encoding, leveraging VisionTransformer to generate positive samples, and applying a multi-clustering voting mechanism to address data labeling and domain shift issues. Experimental results demonstrate that DECMCV achieves classification accuracies of 93.78% and 98.29% on the TrashNet and Huawei Cloud datasets, respectively, outperforming or matching supervised models. On a real-world dataset of 4,169 waste images, only 50 labeled samples were needed to accurately label thousands, improving classification accuracy by 29.85% compared to supervised models. This method effectively addresses style differences, enhances model generalization, and contributes to the advancement of automated waste classification.
The ever-growing availability of unlabeled data presents both opportunities and challenges for training artificial intelligence systems. While self-supervised learning (SSL) has emerged as a powerful paradigm for extracting meaningful representations from vast amounts of unlabeled data, existing methods still struggle to adapt to the non-stationary, non-IID nature of real-world data streams without forgetting previously learned knowledge. Recent works have adopted a repeated cosine annealing schedule for large-scale continual pre-training; however, these schedules (1) inherently cause forgetting during the re-warming phase and (2) have not been systematically compared to existing continual SSL methods. In this work, we systematically compare the widely used cosine schedule with the recently proposed infinite learning rate schedule and empirically find the latter to be a more effective alternative. Our extensive empirical evaluation across diverse image and language datasets demonstrates that the infinite learning rate schedule consistently enhances continual pre-training performance compared to a repeated cosine decay without being restricted to a fixed iteration budget. For instance, in a small-scale MAE pre-training setup, it outperforms several strong baselines from the literature. We then scale up our experiments to larger MAE pre-training and autoregressive language model pre-training. Our results show that the infinite learning rate schedule remains effective at scale, surpassing repeated cosine decay for both MAE pre-training and zero-shot LM benchmarks.
In this paper, we explore a self-supervised model that learns to detect the symmetry of a single object without requiring a dataset-relying solely on the input object itself. We hypothesize that the symmetry of an object can be determined by its intrinsic features, eliminating the need for large datasets during training. Additionally, we design a self-supervised learning strategy that removes the necessity of ground truth labels. These two key elements make our approach both effective and efficient, addressing the prohibitive costs associated with constructing large, labeled datasets for this task. The novelty of our method lies in computing features for each point on the object based on the idea that symmetric points should exhibit similar visual appearances. To achieve this, we leverage features extracted from a foundational image model to compute a visual descriptor for the points. This approach equips the point cloud with visual features that facilitate the optimization of our self-supervised model. Experimental results demonstrate that our method surpasses the state-of-the-art models trained on large datasets. Furthermore, our model is more efficient, effective, and operates with minimal computational and data resources.