State-of-the-art machine learning models, and especially deep learning ones, are significantly data-hungry; they require vast amounts of manually labeled samples to function correctly. However, in most medical imaging fields, obtaining said data can be challenging. Not only the volume of data is a problem, but also the imbalances within its classes; it is common to have many more images of healthy patients than of those with pathology. Computer-aided diagnostic systems suffer from these issues, usually over-designing their models to perform accurately. This work proposes using self-supervised learning for wireless endoscopy videos by introducing a custom-tailored method that does not initially need labels or appropriate balance. We prove that using the inferred inherent structure learned by our method, extracted from the temporal axis, improves the detection rate on several domain-specific applications even under severe imbalance.
Though offering amazing contextualized token-level representations, current pre-trained language models actually take less attention on acquiring sentence-level representation during its self-supervised pre-training. If self-supervised learning can be distinguished into two subcategories, generative and contrastive, then most existing studies show that sentence representation learning may more benefit from the contrastive methods but not the generative methods. However, contrastive learning cannot be well compatible with the common token-level generative self-supervised learning, and does not guarantee good performance on downstream semantic retrieval tasks. Thus, to alleviate such obvious inconveniences, we instead propose a novel generative self-supervised learning objective based on phrase reconstruction. Empirical studies show that our generative learning may yield powerful enough sentence representation and achieve performance in Sentence Textual Similarity (STS) tasks on par with contrastive learning. Further, in terms of unsupervised setting, our generative method outperforms previous state-of-the-art SimCSE on the benchmark of downstream semantic retrieval tasks.
Self-supervised learning has proved to be a powerful approach to learn image representations without the need of large labeled datasets. For underwater robotics, it is of great interest to design computer vision algorithms to improve perception capabilities such as sonar image classification. Due to the confidential nature of sonar imaging and the difficulty to interpret sonar images, it is challenging to create public large labeled sonar datasets to train supervised learning algorithms. In this work, we investigate the potential of three self-supervised learning methods (RotNet, Denoising Autoencoders, and Jigsaw) to learn high-quality sonar image representation without the need of human labels. We present pre-training and transfer learning results on real-life sonar image datasets. Our results indicate that self-supervised pre-training yields classification performance comparable to supervised pre-training in a few-shot transfer learning setup across all three methods. Code and self-supervised pre-trained models are be available at https://github.com/agrija9/ssl-sonar-images
Self-supervised learning in speech involves training a speech representation network on a large-scale unannotated speech corpus, and then applying the learned representations to downstream tasks. Since the majority of the downstream tasks of SSL learning in speech largely focus on the content information in speech, the most desirable speech representations should be able to disentangle unwanted variations, such as speaker variations, from the content. However, disentangling speakers is very challenging, because removing the speaker information could easily result in a loss of content as well, and the damage of the latter usually far outweighs the benefit of the former. In this paper, we propose a new SSL method that can achieve speaker disentanglement without severe loss of content. Our approach is adapted from the HuBERT framework, and incorporates disentangling mechanisms to regularize both the teacher labels and the learned representations. We evaluate the benefit of speaker disentanglement on a set of content-related downstream tasks, and observe a consistent and notable performance advantage of our speaker-disentangled representations.
Rumor detection has become an emerging and active research field in recent years. At the core is to model the rumor characteristics inherent in rich information, such as propagation patterns in social network and semantic patterns in post content, and differentiate them from the truth. However, existing works on rumor detection fall short in modeling heterogeneous information, either using one single information source only (e.g. social network, or post content) or ignoring the relations among multiple sources (e.g. fusing social and content features via simple concatenation). Therefore, they possibly have drawbacks in comprehensively understanding the rumors, and detecting them accurately. In this work, we explore contrastive self-supervised learning on heterogeneous information sources, so as to reveal their relations and characterize rumors better. Technically, we supplement the main supervised task of detection with an auxiliary self-supervised task, which enriches post representations via post self-discrimination. Specifically, given two heterogeneous views of a post (i.e. representations encoding social patterns and semantic patterns), the discrimination is done by maximizing the mutual information between different views of the same post compared to that of other posts. We devise cluster-wise and instance-wise approaches to generate the views and conduct the discrimination, considering different relations of information sources. We term this framework as Self-supervised Rumor Detection (SRD). Extensive experiments on three real-world datasets validate the effectiveness of SRD for automatic rumor detection on social media.
The goal of this paper is to bypass the need for labelled examples in few-shot video understanding at run time. While proven effective, in many practical video settings even labelling a few examples appears unrealistic. This is especially true as the level of details in spatio-temporal video understanding and with it, the complexity of annotations continues to increase. Rather than performing few-shot learning with a human oracle to provide a few densely labelled support videos, we propose to automatically learn to find appropriate support videos given a query. We call this self-shot learning and we outline a simple self-supervised learning method to generate an embedding space well-suited for unsupervised retrieval of relevant samples. To showcase this novel setting, we tackle, for the first time, video instance segmentation in a self-shot (and few-shot) setting, where the goal is to segment instances at the pixel-level across the spatial and temporal domains. We provide strong baseline performances that utilize a novel transformer-based model and show that self-shot learning can even surpass few-shot and can be positively combined for further performance gains. Experiments on new benchmarks show that our approach achieves strong performance, is competitive to oracle support in some settings, scales to large unlabelled video collections, and can be combined in a semi-supervised setting.
Detecting robust keypoints from an image is an integral part of many computer vision problems, and the characteristic orientation and scale of keypoints play an important role for keypoint description and matching. Existing learning-based methods for keypoint detection rely on standard translation-equivariant CNNs but often fail to detect reliable keypoints against geometric variations. To learn to detect robust oriented keypoints, we introduce a self-supervised learning framework using rotation-equivariant CNNs. We propose a dense orientation alignment loss by an image pair generated by synthetic transformations for training a histogram-based orientation map. Our method outperforms the previous methods on an image matching benchmark and a camera pose estimation benchmark.
Crime has become a major concern in many cities, which calls for the rising demand for timely predicting citywide crime occurrence. Accurate crime prediction results are vital for the beforehand decision-making of government to alleviate the increasing concern about the public safety. While many efforts have been devoted to proposing various spatial-temporal forecasting techniques to explore dependence across locations and time periods, most of them follow a supervised learning manner, which limits their spatial-temporal representation ability on sparse crime data. Inspired by the recent success in self-supervised learning, this work proposes a Spatial-Temporal Hypergraph Self-Supervised Learning framework (ST-HSL) to tackle the label scarcity issue in crime prediction. Specifically, we propose the cross-region hypergraph structure learning to encode region-wise crime dependency under the entire urban space. Furthermore, we design the dual-stage self-supervised learning paradigm, to not only jointly capture local- and global-level spatial-temporal crime patterns, but also supplement the sparse crime representation by augmenting region self-discrimination. We perform extensive experiments on two real-life crime datasets. Evaluation results show that our ST-HSL significantly outperforms state-of-the-art baselines. Further analysis provides insights into the superiority of our ST-HSL method in the representation of spatial-temporal crime patterns. The implementation code is available at https://github.com/LZH-YS1998/STHSL.
Point clouds upsampling is a challenging issue to generate dense and uniform point clouds from the given sparse input. Most existing methods either take the end-to-end supervised learning based manner, where large amounts of pairs of sparse input and dense ground-truth are exploited as supervision information; or treat up-scaling of different scale factors as independent tasks, and have to build multiple networks to handle upsampling with varying factors. In this paper, we propose a novel approach that achieves self-supervised and magnification-flexible point clouds upsampling simultaneously. We formulate point clouds upsampling as the task of seeking nearest projection points on the implicit surface for seed points. To this end, we define two implicit neural functions to estimate projection direction and distance respectively, which can be trained by two pretext learning tasks. Experimental results demonstrate that our self-supervised learning based scheme achieves competitive or even better performance than supervised learning based state-of-the-art methods. The source code is publicly available at https://github.com/xnowbzhao/sapcu.
Pre-trained language models (PLMs) like BERT have made significant progress in various downstream NLP tasks. However, by asking models to do cloze-style tests, recent work finds that PLMs are short in acquiring knowledge from unstructured text. To understand the internal behaviour of PLMs in retrieving knowledge, we first define knowledge-baring (K-B) tokens and knowledge-free (K-F) tokens for unstructured text and ask professional annotators to label some samples manually. Then, we find that PLMs are more likely to give wrong predictions on K-B tokens and attend less attention to those tokens inside the self-attention module. Based on these observations, we develop two solutions to help the model learn more knowledge from unstructured text in a fully self-supervised manner. Experiments on knowledge-intensive tasks show the effectiveness of the proposed methods. To our best knowledge, we are the first to explore fully self-supervised learning of knowledge in continual pre-training.
Pre-trained models are essential as feature extractors in modern machine learning systems in various domains. In this study, we hypothesize that representations effective for general audio tasks should provide multiple aspects of robust features of the input sound. For recognizing sounds regardless of perturbations such as varying pitch or timbre, features should be robust to these perturbations. For serving the diverse needs of tasks such as recognition of emotions or music genres, representations should provide multiple aspects of information, such as local and global features. To implement our principle, we propose a self-supervised learning method: Bootstrap Your Own Latent (BYOL) for Audio (BYOL-A, pronounced "viola"). BYOL-A pre-trains representations of the input sound invariant to audio data augmentations, which makes the learned representations robust to the perturbations of sounds. Whereas the BYOL-A encoder combines local and global features and calculates their statistics to make the representation provide multi-aspect information. As a result, the learned representations should provide robust and multi-aspect information to serve various needs of diverse tasks. We evaluated the general audio task performance of BYOL-A compared to previous state-of-the-art methods, and BYOL-A demonstrated generalizability with the best average result of 72.4% and the best VoxCeleb1 result of 57.6%. Extensive ablation experiments revealed that the BYOL-A encoder architecture contributes to most performance, and the final critical portion resorts to the BYOL framework and BYOL-A augmentations. Our code is available online at https://github.com/nttcslab/byol-a for future studies.
Image Captioning (IC) has achieved astonishing developments by incorporating various techniques into the CNN-RNN encoder-decoder architecture. However, since CNN and RNN do not share the basic network component, such a heterogeneous pipeline is hard to be trained end-to-end where the visual encoder will not learn anything from the caption supervision. This drawback inspires the researchers to develop a homogeneous architecture that facilitates end-to-end training, for which Transformer is the perfect one that has proven its huge potential in both vision and language domains and thus can be used as the basic component of the visual encoder and language decoder in an IC pipeline. Meantime, self-supervised learning releases the power of the Transformer architecture that a pre-trained large-scale one can be generalized to various tasks including IC. The success of these large-scale models seems to weaken the importance of the single IC task. However, we demonstrate that IC still has its specific significance in this age by analyzing the connections between IC with some popular self-supervised learning paradigms. Due to the page limitation, we only refer to highly important papers in this short survey and more related works can be found at https://github.com/SjokerLily/awesome-image-captioning.
Recently, self-supervised instance discrimination methods have achieved significant success in learning visual representations from unlabeled photographic images. However, given the marked differences between photographic and medical images, the efficacy of instance-based objectives, focusing on learning the most discriminative global features in the image (i.e., wheels in bicycle), remains unknown in medical imaging. Our preliminary analysis showed that high global similarity of medical images in terms of anatomy hampers instance discrimination methods for capturing a set of distinct features, negatively impacting their performance on medical downstream tasks. To alleviate this limitation, we have developed a simple yet effective self-supervised framework, called Context-Aware instance Discrimination (CAiD). CAiD aims to improve instance discrimination learning by providing finer and more discriminative information encoded from a diverse local context of unlabeled medical images. We conduct a systematic analysis to investigate the utility of the learned features from a three-pronged perspective: (i) generalizability and transferability, (ii) separability in the embedding space, and (iii) reusability. Our extensive experiments demonstrate that CAiD (1) enriches representations learned from existing instance discrimination methods; (2) delivers more discriminative features by adequately capturing finer contextual information from individual medial images; and (3) improves reusability of low/mid-level features compared to standard instance discriminative methods. As open science, all codes and pre-trained models are available on our GitHub page: https://github.com/JLiangLab/CAiD.
Imposing consistency through proxy tasks has been shown to enhance data-driven learning and enable self-supervision in various tasks. This paper introduces novel and effective consistency strategies for optical flow estimation, a problem where labels from real-world data are very challenging to derive. More specifically, we propose occlusion consistency and zero forcing in the forms of self-supervised learning and transformation consistency in the form of semi-supervised learning. We apply these consistency techniques in a way that the network model learns to describe pixel-level motions better while requiring no additional annotations. We demonstrate that our consistency strategies applied to a strong baseline network model using the original datasets and labels provide further improvements, attaining the state-of-the-art results on the KITTI-2015 scene flow benchmark in the non-stereo category. Our method achieves the best foreground accuracy (4.33% in Fl-all) over both the stereo and non-stereo categories, even though using only monocular image inputs.
Performing transductive learning on graphs with very few labeled data, that is, two or three samples for each category, is challenging due to the lack of supervision. In the existing work, self-supervised learning via a single view model is widely adopted to address the problem. However, recent observation shows multiview representations of an object share the same semantic information in high-level feature space. For each sample, we generate heterogeneous representations and use view-consistency loss to make their representations consistent with each other. Multiview representation also inspires to supervise the pseudolabels generation by the aid of mutual supervision between views. In this article, we thus propose a view-consistent heterogeneous network (VCHN) to learn better representations by aligning view-agnostic semantics. Specifically, VCHN is constructed by constraining the predictions between two views so that the view pairs can supervise each other. To make the best use of cross-view information, we further propose a novel training strategy to generate more reliable pseudolabels, which thus enhances predictions of the VCHN. Extensive experimental results on three benchmark datasets demonstrate that our method achieves superior performance over state-of-the-art methods under very low label rates.
A Vision Transformer (ViT) is a simple neural architecture amenable to serve several computer vision tasks. It has limited built-in architectural priors, in contrast to more recent architectures that incorporate priors either about the input data or of specific tasks. Recent works show that ViTs benefit from self-supervised pre-training, in particular BerT-like pre-training like BeiT. In this paper, we revisit the supervised training of ViTs. Our procedure builds upon and simplifies a recipe introduced for training ResNet-50. It includes a new simple data-augmentation procedure with only 3 augmentations, closer to the practice in self-supervised learning. Our evaluations on Image classification (ImageNet-1k with and without pre-training on ImageNet-21k), transfer learning and semantic segmentation show that our procedure outperforms by a large margin previous fully supervised training recipes for ViT. It also reveals that the performance of our ViT trained with supervision is comparable to that of more recent architectures. Our results could serve as better baselines for recent self-supervised approaches demonstrated on ViT.
We propose Masked Siamese Networks (MSN), a self-supervised learning framework for learning image representations. Our approach matches the representation of an image view containing randomly masked patches to the representation of the original unmasked image. This self-supervised pre-training strategy is particularly scalable when applied to Vision Transformers since only the unmasked patches are processed by the network. As a result, MSNs improve the scalability of joint-embedding architectures, while producing representations of a high semantic level that perform competitively on low-shot image classification. For instance, on ImageNet-1K, with only 5,000 annotated images, our base MSN model achieves 72.4% top-1 accuracy, and with 1% of ImageNet-1K labels, we achieve 75.7% top-1 accuracy, setting a new state-of-the-art for self-supervised learning on this benchmark. Our code is publicly available.
Despite recent advancements in deep learning technologies, Child Speech Recognition remains a challenging task. Current Automatic Speech Recognition (ASR) models require substantial amounts of annotated data for training, which is scarce. In this work, we explore using the ASR model, wav2vec2, with different pretraining and finetuning configurations for self-supervised learning (SSL) toward improving automatic child speech recognition. The pretrained wav2vec2 models were finetuned using different amounts of child speech training data, adult speech data, and a combination of both, to discover the optimum amount of data required to finetune the model for the task of child ASR. Our trained model achieves the best Word Error Rate (WER) of 7.42 on the MyST child speech dataset, 2.99 on the PFSTAR dataset and 12.47 on the CMU KIDS dataset as compared to any other previous methods. Our models outperformed the wav2vec2 BASE 960 on child speech which is considered a state-of-the-art ASR model on adult speech by just using 10 hours of child speech data in finetuning. The analysis of different types of training data and their effect on inference is also provided by using a combination of datasets in pretraining, finetuning and inference.
Self-supervised learning (SSL) has attracted much interest in remote sensing and earth observation due to its ability to learn task-agnostic representations without human annotation. While most of the existing SSL works in remote sensing utilize ConvNet backbones and focus on a single modality, we explore the potential of vision transformers (ViTs) for joint SAR-optical representation learning. Based on DINO, a state-of-the-art SSL algorithm that distills knowledge from two augmented views of an input image, we combine SAR and optical imagery by concatenating all channels to a unified input. Subsequently, we randomly mask out channels of one modality as a data augmentation strategy. While training, the model gets fed optical-only, SAR-only, and SAR-optical image pairs learning both inner- and intra-modality representations. Experimental results employing the BigEarthNet-MM dataset demonstrate the benefits of both, the ViT backbones and the proposed multimodal SSL algorithm DINO-MM.
User sessions empower many search and recommendation tasks on a daily basis. Such session data are semi-structured, which encode heterogeneous relations between queries and products, and each item is described by the unstructured text. Despite recent advances in self-supervised learning for text or graphs, there lack of self-supervised learning models that can effectively capture both intra-item semantics and inter-item interactions for semi-structured sessions. To fill this gap, we propose CERES, a graph-based transformer model for semi-structured session data. CERES learns representations that capture both inter- and intra-item semantics with (1) a graph-conditioned masked language pretraining task that jointly learns from item text and item-item relations; and (2) a graph-conditioned transformer architecture that propagates inter-item contexts to item-level representations. We pretrained CERES using ~468 million Amazon sessions and find that CERES outperforms strong pretraining baselines by up to 9% in three session search and entity linking tasks.