This paper introduces vox2vec - a contrastive method for self-supervised learning (SSL) of voxel-level representations. vox2vec representations are modeled by a Feature Pyramid Network (FPN): a voxel representation is a concatenation of the corresponding feature vectors from different pyramid levels. The FPN is pre-trained to produce similar representations for the same voxel in different augmented contexts and distinctive representations for different voxels. This results in unified multi-scale representations that capture both global semantics (e.g., body part) and local semantics (e.g., different small organs or healthy versus tumor tissue). We use vox2vec to pre-train a FPN on more than 6500 publicly available computed tomography images. We evaluate the pre-trained representations by attaching simple heads on top of them and training the resulting models for 22 segmentation tasks. We show that vox2vec outperforms existing medical imaging SSL techniques in three evaluation setups: linear and non-linear probing and end-to-end fine-tuning. Moreover, a non-linear head trained on top of the frozen vox2vec representations achieves competitive performance with the FPN trained from scratch while having 50 times fewer trainable parameters. The code is available at https://github.com/mishgon/vox2vec .
Gait, the manner of walking, has been proven to be a reliable biometric with uses in surveillance, marketing and security. A promising new direction for the field is training gait recognition systems without explicit human annotations, through self-supervised learning approaches. Such methods are heavily reliant on strong augmentations for the same walking sequence to induce more data variability and to simulate additional walking variations. Current data augmentation schemes are heuristic and cannot provide the necessary data variation as they are only able to provide simple temporal and spatial distortions. In this work, we propose GaitMorph, a novel method to modify the walking variation for an input gait sequence. Our method entails the training of a high-compression model for gait skeleton sequences that leverages unlabelled data to construct a discrete and interpretable latent space, which preserves identity-related features. Furthermore, we propose a method based on optimal transport theory to learn latent transport maps on the discrete codebook that morph gait sequences between variations. We perform extensive experiments and show that our method is suitable to synthesize additional views for an input sequence.
Self-supervised learning is popular method because of its ability to learn features in images without using its labels and is able to overcome limited labeled datasets used in supervised learning. Self-supervised learning works by using a pretext task which will be trained on the model before being applied to a specific task. There are some examples of pretext tasks used in self-supervised learning in the field of image recognition, namely rotation prediction, solving jigsaw puzzles, and predicting relative positions on image. Previous studies have only used one type of transformation as a pretext task. This raises the question of how it affects if more than one pretext task is used and to use a gating network to combine all pretext tasks. Therefore, we propose the Gated Self-Supervised Learning method to improve image classification which use more than one transformation as pretext task and uses the Mixture of Expert architecture as a gating network in combining each pretext task so that the model automatically can study and focus more on the most useful augmentations for classification. We test performance of the proposed method in several scenarios, namely CIFAR imbalance dataset classification, adversarial perturbations, Tiny-Imagenet dataset classification, and semi-supervised learning. Moreover, there are Grad-CAM and T-SNE analysis that are used to see the proposed method for identifying important features that influence image classification and representing data for each class and separating different classes properly. Our code is in https://github.com/aristorenaldo/G-SSL
Fluorescent Neuronal Cells v2 is a collection of fluorescence microscopy images and the corresponding ground-truth annotations, designed to foster innovative research in the domains of Life Sciences and Deep Learning. This dataset encompasses three image collections in which rodent neuronal cells' nuclei and cytoplasm are stained with diverse markers to highlight their anatomical or functional characteristics. Alongside the images, we provide ground-truth annotations for several learning tasks, including semantic segmentation, object detection, and counting. The contribution is two-fold. First, given the variety of annotations and their accessible formats, we envision our work facilitating methodological advancements in computer vision approaches for segmentation, detection, feature learning, unsupervised and self-supervised learning, transfer learning, and related areas. Second, by enabling extensive exploration and benchmarking, we hope Fluorescent Neuronal Cells v2 will catalyze breakthroughs in fluorescence microscopy analysis and promote cutting-edge discoveries in life sciences. The data are available at: https://amsacta.unibo.it/id/eprint/7347
Preserving training dynamics across batch sizes is an important tool for practical machine learning as it enables the trade-off between batch size and wall-clock time. This trade-off is typically enabled by a scaling rule, for example, in stochastic gradient descent, one should scale the learning rate linearly with the batch size. Another important machine learning tool is the model EMA, a functional copy of a target model, whose parameters move towards those of its target model according to an Exponential Moving Average (EMA) at a rate parameterized by a momentum hyperparameter. This model EMA can improve the robustness and generalization of supervised learning, stabilize pseudo-labeling, and provide a learning signal for Self-Supervised Learning (SSL). Prior works have not considered the optimization of the model EMA when performing scaling, leading to different training dynamics across batch sizes and lower model performance. In this work, we provide a scaling rule for optimization in the presence of a model EMA and demonstrate the rule's validity across a range of architectures, optimizers, and data modalities. We also show the rule's validity where the model EMA contributes to the optimization of the target model, enabling us to train EMA-based pseudo-labeling and SSL methods at small and large batch sizes. For SSL, we enable training of BYOL up to batch size 24,576 without sacrificing performance, a 6$\times$ wall-clock time reduction under idealized hardware settings.
Emotion recognition models using audio input data can enable the development of interactive systems with applications in mental healthcare, marketing, gaming, and social media analysis. While the field of affective computing using audio data is rich, a major barrier to achieve consistently high-performance models is the paucity of available training labels. Self-supervised learning (SSL) is a family of methods which can learn despite a scarcity of supervised labels by predicting properties of the data itself. To understand the utility of self-supervised learning for audio-based emotion recognition, we have applied self-supervised learning pre-training to the classification of emotions from the CMU- MOSEI's acoustic modality. Unlike prior papers that have experimented with raw acoustic data, our technique has been applied to encoded acoustic data. Our model is first pretrained to uncover the randomly-masked timestamps of the acoustic data. The pre-trained model is then fine-tuned using a small sample of annotated data. The performance of the final model is then evaluated via several evaluation metrics against a baseline deep learning model with an identical backbone architecture. We find that self-supervised learning consistently improves the performance of the model across all metrics. This work shows the utility of self-supervised learning for affective computing, demonstrating that self-supervised learning is most useful when the number of training examples is small, and that the effect is most pronounced for emotions which are easier to classify such as happy, sad and anger. This work further demonstrates that self-supervised learning works when applied to embedded feature representations rather than the traditional approach of pre-training on the raw input space.
Neural speech separation has made remarkable progress and its integration with automatic speech recognition (ASR) is an important direction towards realizing multi-speaker ASR. This work provides an insightful investigation of speech separation in reverberant and noisy-reverberant scenarios as an ASR front-end. In detail, we explore multi-channel separation methods, mask-based beamforming and complex spectral mapping, as well as the best features to use in the ASR back-end model. We employ the recent self-supervised learning representation (SSLR) as a feature and improve the recognition performance from the case with filterbank features. To further improve multi-speaker recognition performance, we present a carefully designed training strategy for integrating speech separation and recognition with SSLR. The proposed integration using TF-GridNet-based complex spectral mapping and WavLM-based SSLR achieves a 2.5% word error rate in reverberant WHAMR! test set, significantly outperforming an existing mask-based MVDR beamforming and filterbank integration (28.9%).
Being able to perceive the semantics and the spatial structure of the environment is essential for visual navigation of a household robot. However, most existing works only employ visual backbones pre-trained either with independent images for classification or with self-supervised learning methods to adapt to the indoor navigation domain, neglecting the spatial relationships that are essential to the learning of navigation. Inspired by the behavior that humans naturally build semantically and spatially meaningful cognitive maps in their brains during navigation, in this paper, we propose a novel navigational-specific visual representation learning method by contrasting the agent's egocentric views and semantic maps (Ego$^2$-Map). We apply the visual transformer as the backbone encoder and train the model with data collected from the large-scale Habitat-Matterport3D environments. Ego$^2$-Map learning transfers the compact and rich information from a map, such as objects, structure and transition, to the agent's egocentric representations for navigation. Experiments show that agents using our learned representations on object-goal navigation outperform recent visual pre-training methods. Moreover, our representations significantly improve vision-and-language navigation in continuous environments for both high-level and low-level action spaces, achieving new state-of-the-art results of 47% SR and 41% SPL on the test server.
Self-supervised learning usually uses a large amount of unlabeled data to pre-train an encoder which can be used as a general-purpose feature extractor, such that downstream users only need to perform fine-tuning operations to enjoy the benefit of "large model". Despite this promising prospect, the security of pre-trained encoder has not been thoroughly investigated yet, especially when the pre-trained encoder is publicly available for commercial use. In this paper, we propose AdvEncoder, the first framework for generating downstream-agnostic universal adversarial examples based on the pre-trained encoder. AdvEncoder aims to construct a universal adversarial perturbation or patch for a set of natural images that can fool all the downstream tasks inheriting the victim pre-trained encoder. Unlike traditional adversarial example works, the pre-trained encoder only outputs feature vectors rather than classification labels. Therefore, we first exploit the high frequency component information of the image to guide the generation of adversarial examples. Then we design a generative attack framework to construct adversarial perturbations/patches by learning the distribution of the attack surrogate dataset to improve their attack success rates and transferability. Our results show that an attacker can successfully attack downstream tasks without knowing either the pre-training dataset or the downstream dataset. We also tailor four defenses for pre-trained encoders, the results of which further prove the attack ability of AdvEncoder.
Urban modeling from LiDAR point clouds is an important topic in computer vision, computer graphics, photogrammetry and remote sensing. 3D city models have found a wide range of applications in smart cities, autonomous navigation, urban planning and mapping etc. However, existing datasets for 3D modeling mainly focus on common objects such as furniture or cars. Lack of building datasets has become a major obstacle for applying deep learning technology to specific domains such as urban modeling. In this paper, we present a urban-scale dataset consisting of more than 160 thousands buildings along with corresponding point clouds, mesh and wire-frame models, covering 16 cities in Estonia about 998 Km2. We extensively evaluate performance of state-of-the-art algorithms including handcrafted and deep feature based methods. Experimental results indicate that Building3D has challenges of high intra-class variance, data imbalance and large-scale noises. The Building3D is the first and largest urban-scale building modeling benchmark, allowing a comparison of supervised and self-supervised learning methods. We believe that our Building3D will facilitate future research on urban modeling, aerial path planning, mesh simplification, and semantic/part segmentation etc.
Recent advancements in large-scale Vision Transformers have made significant strides in improving pre-trained models for medical image segmentation. However, these methods face a notable challenge in acquiring a substantial amount of pre-training data, particularly within the medical field. To address this limitation, we present Masked Multi-view with Swin Transformers (SwinMM), a novel multi-view pipeline for enabling accurate and data-efficient self-supervised medical image analysis. Our strategy harnesses the potential of multi-view information by incorporating two principal components. In the pre-training phase, we deploy a masked multi-view encoder devised to concurrently train masked multi-view observations through a range of diverse proxy tasks. These tasks span image reconstruction, rotation, contrastive learning, and a novel task that employs a mutual learning paradigm. This new task capitalizes on the consistency between predictions from various perspectives, enabling the extraction of hidden multi-view information from 3D medical data. In the fine-tuning stage, a cross-view decoder is developed to aggregate the multi-view information through a cross-attention block. Compared with the previous state-of-the-art self-supervised learning method Swin UNETR, SwinMM demonstrates a notable advantage on several medical image segmentation tasks. It allows for a smooth integration of multi-view information, significantly boosting both the accuracy and data-efficiency of the model. Code and models are available at https://github.com/UCSC-VLAA/SwinMM/.
Contrastive learning models based on Siamese structure have demonstrated remarkable performance in self-supervised learning. Such a success of contrastive learning relies on two conditions, a sufficient number of positive pairs and adequate variations between them. If the conditions are not met, these frameworks will lack semantic contrast and be fragile on overfitting. To address these two issues, we propose Hallucinator that could efficiently generate additional positive samples for further contrast. The Hallucinator is differentiable and creates new data in the feature space. Thus, it is optimized directly with the pre-training task and introduces nearly negligible computation. Moreover, we reduce the mutual information of hallucinated pairs and smooth them through non-linear operations. This process helps avoid over-confident contrastive learning models during the training and achieves more transformation-invariant feature embeddings. Remarkably, we empirically prove that the proposed Hallucinator generalizes well to various contrastive learning models, including MoCoV1&V2, SimCLR and SimSiam. Under the linear classification protocol, a stable accuracy gain is achieved, ranging from 0.3% to 3.0% on CIFAR10&100, Tiny ImageNet, STL-10 and ImageNet. The improvement is also observed in transferring pre-train encoders to the downstream tasks, including object detection and segmentation.
Depth estimation is a crucial task for the creation of depth maps, one of the most important components for augmented reality (AR) and other applications. However, the most widely used hardware for AR and smartphones has only sparse depth sensors with different ground truth depth acquisition methods. Thus, depth estimation models that are robust for downstream AR tasks performance can only be trained reliably using self-supervised learning based on camera information. Previous works in the field mostly focus on self-supervised models with pure convolutional architectures, without taking global spatial context into account.In this paper, we utilize vision transformer architectures for self-supervised monocular depth estimation and propose VTDepth, a vision transformer-based model, which provides a solution to the problem of the global spatial context. We compare various combinations of convolutional and transformer architectures for self-supervised depth estimation and show that the best combination of models is an encoder with a transformer basis and convolutional decoder. Our experiments demonstrate the efficiency of VTDepth for self-supervised depth estimation. Our set of models achieves state-of-the-art performance for self-supervised learning on NYUv2 and KITTI datasets. Our code is available at https://github.com/ahbpp/VTDepth.
Self-supervised learning of visual representations has been focusing on learning content features, which do not capture object motion or location, and focus on identifying and differentiating objects in images and videos. On the other hand, optical flow estimation is a task that does not involve understanding the content of the images on which it is estimated. We unify the two approaches and introduce MC-JEPA, a joint-embedding predictive architecture and self-supervised learning approach to jointly learn optical flow and content features within a shared encoder, demonstrating that the two associated objectives; the optical flow estimation objective and the self-supervised learning objective; benefit from each other and thus learn content features that incorporate motion information. The proposed approach achieves performance on-par with existing unsupervised optical flow benchmarks, as well as with common self-supervised learning approaches on downstream tasks such as semantic segmentation of images and videos.
Handwriting authentication is a valuable tool used in various fields, such as fraud prevention and cultural heritage protection. However, it remains a challenging task due to the complex features, severe damage, and lack of supervision. In this paper, we propose a novel Contrastive Self-Supervised Learning framework for Robust Handwriting Authentication (CSSL-RHA) to address these issues. It can dynamically learn complex yet important features and accurately predict writer identities. Specifically, to remove the negative effects of imperfections and redundancy, we design an information-theoretic filter for pre-processing and propose a novel adaptive matching scheme to represent images as patches of local regions dominated by more important features. Through online optimization at inference time, the most informative patch embeddings are identified as the "most important" elements. Furthermore, we employ contrastive self-supervised training with a momentum-based paradigm to learn more general statistical structures of handwritten data without supervision. We conduct extensive experiments on five benchmark datasets and our manually annotated dataset EN-HA, which demonstrate the superiority of our CSSL-RHA compared to baselines. Additionally, we show that our proposed model can still effectively achieve authentication even under abnormal circumstances, such as data falsification and corruption.
In the recent years, speech representation learning is constructed primarily as a self-supervised learning (SSL) task, using the raw audio signal alone, while ignoring the side-information that is often available for a given speech recording. In this paper, we propose MASR, a Multi-label Aware Speech Representation learning framework, which addresses the aforementioned limitations. MASR enables the inclusion of multiple external knowledge sources to enhance the utilization of meta-data information. The external knowledge sources are incorporated in the form of sample-level pair-wise similarity matrices that are useful in a hard-mining loss. A key advantage of the MASR framework is that it can be combined with any choice of SSL method. Using MASR representations, we perform evaluations on several downstream tasks such as language identification, speech recognition and other non-semantic tasks such as speaker and emotion recognition. In these experiments, we illustrate significant performance improvements for the MASR over other established benchmarks. We perform a detailed analysis on the language identification task to provide insights on how the proposed loss function enables the representations to separate closely related languages.
Self-supervised learning (SSL) for clinical time series data has received significant attention in recent literature, since these data are highly rich and provide important information about a patient's physiological state. However, most existing SSL methods for clinical time series are limited in that they are designed for unimodal time series, such as a sequence of structured features (e.g., lab values and vitals signs) or an individual high-dimensional physiological signal (e.g., an electrocardiogram). These existing methods cannot be readily extended to model time series that exhibit multimodality, with structured features and high-dimensional data being recorded at each timestep in the sequence. In this work, we address this gap and propose a new SSL method -- Sequential Multi-Dimensional SSL -- where a SSL loss is applied both at the level of the entire sequence and at the level of the individual high-dimensional data points in the sequence in order to better capture information at both scales. Our strategy is agnostic to the specific form of loss function used at each level -- it can be contrastive, as in SimCLR, or non-contrastive, as in VICReg. We evaluate our method on two real-world clinical datasets, where the time series contains sequences of (1) high-frequency electrocardiograms and (2) structured data from lab values and vitals signs. Our experimental results indicate that pre-training with our method and then fine-tuning on downstream tasks improves performance over baselines on both datasets, and in several settings, can lead to improvements across different self-supervised loss functions.
Recent contrastive language image pre-training has led to learning highly transferable and robust image representations. However, adapting these models to video domains with minimal supervision remains an open problem. We explore a simple step in that direction, using language tied self-supervised learning to adapt an image CLIP model to the video domain. A backbone modified for temporal modeling is trained under self-distillation settings with train objectives operating in an action concept space. Feature vectors of various action concepts extracted from a language encoder using relevant textual prompts construct this space. We introduce two train objectives, concept distillation and concept alignment, that retain generality of original representations while enforcing relations between actions and their attributes. Our approach improves zero-shot and linear probing performance on three action recognition benchmarks.
Despite the rapid progress in self-supervised learning (SSL), end-to-end fine-tuning still remains the dominant fine-tuning strategy for medical imaging analysis. However, it remains unclear whether this approach is truly optimal for effectively utilizing the pre-trained knowledge, especially considering the diverse categories of SSL that capture different types of features. In this paper, we present the first comprehensive study that discovers effective fine-tuning strategies for self-supervised learning in medical imaging. After developing strong contrastive and restorative SSL baselines that outperform SOTA methods across four diverse downstream tasks, we conduct an extensive fine-tuning analysis across multiple pre-training and fine-tuning datasets, as well as various fine-tuning dataset sizes. Contrary to the conventional wisdom of fine-tuning only the last few layers of a pre-trained network, we show that fine-tuning intermediate layers is more effective, with fine-tuning the second quarter (25-50%) of the network being optimal for contrastive SSL whereas fine-tuning the third quarter (50-75%) of the network being optimal for restorative SSL. Compared to the de-facto standard of end-to-end fine-tuning, our best fine-tuning strategy, which fine-tunes a shallower network consisting of the first three quarters (0-75%) of the pre-trained network, yields improvements of as much as 5.48%. Additionally, using these insights, we propose a simple yet effective method to leverage the complementary strengths of multiple SSL models, resulting in enhancements of up to 3.57% compared to using the best model alone. Hence, our fine-tuning strategies not only enhance the performance of individual SSL models, but also enable effective utilization of the complementary strengths offered by multiple SSL models, leading to significant improvements in self-supervised medical imaging analysis.
The mechanisms behind the success of multi-view self-supervised learning (MVSSL) are not yet fully understood. Contrastive MVSSL methods have been studied through the lens of InfoNCE, a lower bound of the Mutual Information (MI). However, the relation between other MVSSL methods and MI remains unclear. We consider a different lower bound on the MI consisting of an entropy and a reconstruction term (ER), and analyze the main MVSSL families through its lens. Through this ER bound, we show that clustering-based methods such as DeepCluster and SwAV maximize the MI. We also re-interpret the mechanisms of distillation-based approaches such as BYOL and DINO, showing that they explicitly maximize the reconstruction term and implicitly encourage a stable entropy, and we confirm this empirically. We show that replacing the objectives of common MVSSL methods with this ER bound achieves competitive performance, while making them stable when training with smaller batch sizes or smaller exponential moving average (EMA) coefficients. Github repo: https://github.com/apple/ml-entropy-reconstruction.