Whilst computer vision models built using self-supervised approaches are now commonplace, some important questions remain. Do self-supervised models learn highly redundant channel features? What if a self-supervised network could dynamically select the important channels and get rid of the unnecessary ones? Currently, convnets pre-trained with self-supervision have obtained comparable performance on downstream tasks in comparison to their supervised counterparts in computer vision. However, there are drawbacks to self-supervised models including their large numbers of parameters, computationally expensive training strategies and a clear need for faster inference on downstream tasks. In this work, our goal is to address the latter by studying how a standard channel selection method developed for supervised learning can be applied to networks trained with self-supervision. We validate our findings on a range of target budgets $t_{d}$ for channel computation on image classification task across different datasets, specifically CIFAR-10, CIFAR-100, and ImageNet-100, obtaining comparable performance to that of the original network when selecting all channels but at a significant reduction in computation reported in terms of FLOPs.
In this paper, we present end-to-end and speech embedding based systems trained in a self-supervised fashion to participate in the ACM Multimedia 2022 ComParE Challenge, specifically the stuttering sub-challenge. In particular, we exploit the embeddings from the pre-trained Wav2Vec2.0 model for stuttering detection (SD) on the KSoF dataset. After embedding extraction, we benchmark with several methods for SD. Our proposed self-supervised based SD system achieves a UAR of 36.9% and 41.0% on validation and test sets respectively, which is 31.32% (validation set) and 1.49% (test set) higher than the best (DeepSpectrum) challenge baseline (CBL). Moreover, we show that concatenating layer embeddings with Mel-frequency cepstral coefficients (MFCCs) features further improves the UAR of 33.81% and 5.45% on validation and test sets respectively over the CBL. Finally, we demonstrate that the summing information across all the layers of Wav2Vec2.0 surpasses the CBL by a relative margin of 45.91% and 5.69% on validation and test sets respectively. Grand-challenge: Computational Paralinguistics ChallengE
Self-supervised learning has recently achieved great success in representation learning without human annotations. The dominant method -- that is contrastive learning, is generally based on instance discrimination tasks, i.e., individual samples are treated as independent categories. However, presuming all the samples are different contradicts the natural grouping of similar samples in common visual datasets, e.g., multiple views of the same dog. To bridge the gap, this paper proposes an adaptive method that introduces soft inter-sample relations, namely Adaptive Soft Contrastive Learning (ASCL). More specifically, ASCL transforms the original instance discrimination task into a multi-instance soft discrimination task, and adaptively introduces inter-sample relations. As an effective and concise plug-in module for existing self-supervised learning frameworks, ASCL achieves the best performance on several benchmarks in terms of both performance and efficiency. Code is available at https://github.com/MrChenFeng/ASCL_ICPR2022.
Learning representations of neural network weights given a model zoo is an emerging and challenging area with many potential applications from model inspection, to neural architecture search or knowledge distillation. Recently, an autoencoder trained on a model zoo was able to learn a hyper-representation, which captures intrinsic and extrinsic properties of the models in the zoo. In this work, we extend hyper-representations for generative use to sample new model weights as pre-training. We propose layer-wise loss normalization which we demonstrate is key to generate high-performing models and a sampling method based on the empirical density of hyper-representations. The models generated using our methods are diverse, performant and capable to outperform conventional baselines for transfer learning. Our results indicate the potential of knowledge aggregation from model zoos to new models via hyper-representations thereby paving the avenue for novel research directions.
Recent developments in self-supervised learning give us the possibility to further reduce human intervention in multi-step pipelines where the focus evolves around particular objects of interest. In the present paper, the focus lays in the nuclei in histopathology images. In particular we aim at extracting cellular information in an unsupervised manner for a downstream task. As nuclei present themselves in a variety of sizes, we propose a new Scale-dependant convolutional layer to bypass scaling issues when resizing nuclei. On three nuclei datasets, we benchmark the following methods: handcrafted, pre-trained ResNet, supervised ResNet and self-supervised features. We show that the proposed convolution layer boosts performance and that this layer combined with Barlows-Twins allows for better nuclei encoding compared to the supervised paradigm in the low sample setting and outperforms all other proposed unsupervised methods. In addition, we extend the existing TNBC dataset to incorporate nuclei class annotation in order to enrich and publicly release a small sample setting dataset for nuclei segmentation and classification.
Contrastive self-supervised learning (CSL) has managed to match or surpass the performance of supervised learning in image and video classification. However, it is still largely unknown if the nature of the representations induced by the two learning paradigms is similar. We investigate this under the lens of adversarial robustness. Our analysis of the problem reveals that CSL has intrinsically higher sensitivity to perturbations over supervised learning. We identify the uniform distribution of data representation over a unit hypersphere in the CSL representation space as the key contributor to this phenomenon. We establish that this is a result of the presence of false negative pairs in the training process, which increases model sensitivity to input perturbations. Our finding is supported by extensive experiments for image and video classification using adversarial perturbations and other input corruptions. We devise a strategy to detect and remove false negative pairs that is simple, yet effective in improving model robustness with CSL training. We close up to 68% of the robustness gap between CSL and its supervised counterpart. Finally, we contribute to adversarial learning by incorporating our method in CSL. We demonstrate an average gain of about 5% over two different state-of-the-art methods in this domain.
Many successful methods developed for medical image analysis that are based on machine learning use supervised learning approaches, which often require large datasets annotated by experts to achieve high accuracy. However, medical data annotation is time-consuming and expensive, especially for segmentation tasks. To solve the problem of learning with limited labeled medical image data, an alternative deep learning training strategy based on self-supervised pretraining on unlabeled MRI scans is proposed in this work. Our pretraining approach first, randomly applies different distortions to random areas of unlabeled images and then predicts the type of distortions and loss of information. To this aim, an improved version of Mask-RCNN architecture has been adapted to localize the distortion location and recover the original image pixels. The effectiveness of the proposed method for segmentation tasks in different pre-training and fine-tuning scenarios is evaluated based on the Osteoarthritis Initiative dataset. Using this self-supervised pretraining method improved the Dice score by 20% compared to training from scratch. The proposed self-supervised learning is simple, effective, and suitable for different ranges of medical image analysis tasks including anomaly detection, segmentation, and classification.
Adversarial training (AT) for robust representation learning and self-supervised learning (SSL) for unsupervised representation learning are two active research fields. Integrating AT into SSL, multiple prior works have accomplished a highly significant yet challenging task: learning robust representation without labels. A widely used framework is adversarial contrastive learning which couples AT and SSL, and thus constitute a very complex optimization problem. Inspired by the divide-and-conquer philosophy, we conjecture that it might be simplified as well as improved by solving two sub-problems: non-robust SSL and pseudo-supervised AT. This motivation shifts the focus of the task from seeking an optimal integrating strategy for a coupled problem to finding sub-solutions for sub-problems. With this said, this work discards prior practices of directly introducing AT to SSL frameworks and proposed a two-stage framework termed Decoupled Adversarial Contrastive Learning (DeACL). Extensive experimental results demonstrate that our DeACL achieves SOTA self-supervised adversarial robustness while significantly reducing the training time, which validates its effectiveness and efficiency. Moreover, our DeACL constitutes a more explainable solution, and its success also bridges the gap with semi-supervised AT for exploiting unlabeled samples for robust representation learning. The code is publicly accessible at https://github.com/pantheon5100/DeACL.
The hardware challenges associated with light-field(LF) imaging has made it difficult for consumers to access its benefits like applications in post-capture focus and aperture control. Learning-based techniques which solve the ill-posed problem of LF reconstruction from sparse (1, 2 or 4) views have significantly reduced the requirement for complex hardware. LF video reconstruction from sparse views poses a special challenge as acquiring ground-truth for training these models is hard. Hence, we propose a self-supervised learning-based algorithm for LF video reconstruction from monocular videos. We use self-supervised geometric, photometric and temporal consistency constraints inspired from a recent self-supervised technique for LF video reconstruction from stereo video. Additionally, we propose three key techniques that are relevant to our monocular video input. We propose an explicit disocclusion handling technique that encourages the network to inpaint disoccluded regions in a LF frame, using information from adjacent input temporal frames. This is crucial for a self-supervised technique as a single input frame does not contain any information about the disoccluded regions. We also propose an adaptive low-rank representation that provides a significant boost in performance by tailoring the representation to each input scene. Finally, we also propose a novel refinement block that is able to exploit the available LF image data using supervised learning to further refine the reconstruction quality. Our qualitative and quantitative analysis demonstrates the significance of each of the proposed building blocks and also the superior results compared to previous state-of-the-art monocular LF reconstruction techniques. We further validate our algorithm by reconstructing LF videos from monocular videos acquired using a commercial GoPro camera.
Supervised multi-view stereo (MVS) methods have achieved remarkable progress in terms of reconstruction quality, but suffer from the challenge of collecting large-scale ground-truth depth. In this paper, we propose a novel self-supervised training pipeline for MVS based on knowledge distillation, termed KD-MVS, which mainly consists of self-supervised teacher training and distillation-based student training. Specifically, the teacher model is trained in a self-supervised fashion using both photometric and featuremetric consistency. Then we distill the knowledge of the teacher model to the student model through probabilistic knowledge transferring. With the supervision of validated knowledge, the student model is able to outperform its teacher by a large margin. Extensive experiments performed on multiple datasets show our method can even outperform supervised methods.
Video Anomaly Detection (VAD) is an important topic in computer vision. Motivated by the recent advances in self-supervised learning, this paper addresses VAD by solving an intuitive yet challenging pretext task, i.e., spatio-temporal jigsaw puzzles, which is cast as a multi-label fine-grained classification problem. Our method exhibits several advantages over existing works: 1) the spatio-temporal jigsaw puzzles are decoupled in terms of spatial and temporal dimensions, responsible for capturing highly discriminative appearance and motion features, respectively; 2) full permutations are used to provide abundant jigsaw puzzles covering various difficulty levels, allowing the network to distinguish subtle spatio-temporal differences between normal and abnormal events; and 3) the pretext task is tackled in an end-to-end manner without relying on any pre-trained models. Our method outperforms state-of-the-art counterparts on three public benchmarks. Especially on ShanghaiTech Campus, the result is superior to reconstruction and prediction-based methods by a large margin.
Clustering is a ubiquitous tool in unsupervised learning. Most of the existing self-supervised representation learning methods typically cluster samples based on visually dominant features. While this works well for image-based self-supervision, it often fails for videos, which require understanding motion rather than focusing on background. Using optical flow as complementary information to RGB can alleviate this problem. However, we observe that a naive combination of the two views does not provide meaningful gains. In this paper, we propose a principled way to combine two views. Specifically, we propose a novel clustering strategy where we use the initial cluster assignment of each view as prior to guide the final cluster assignment of the other view. This idea will enforce similar cluster structures for both views, and the formed clusters will be semantically abstract and robust to noisy inputs coming from each individual view. Additionally, we propose a novel regularization strategy to address the feature collapse problem, which is common in cluster-based self-supervised learning methods. Our extensive evaluation shows the effectiveness of our learned representations on downstream tasks, e.g., video retrieval and action recognition. Specifically, we outperform the state of the art by 7% on UCF and 4% on HMDB for video retrieval, and 5% on UCF and 6% on HMDB for video classification
Relying on deep supervised or self-supervised learning, previous methods for depth completion from paired single image and sparse depth data have achieved impressive performance in recent years. However, facing a new environment where the test data occurs online and differs from the training data in the RGB image content and depth sparsity, the trained model might suffer severe performance drop. To encourage the trained model to work well in such conditions, we expect it to be capable of adapting to the new environment continuously and effectively. To achieve this, we propose MetaComp. It utilizes the meta-learning technique to simulate adaptation policies during the training phase, and then adapts the model to new environments in a self-supervised manner in testing. Considering that the input is multi-modal data, it would be challenging to adapt a model to variations in two modalities simultaneously, due to significant differences in structure and form of the two modal data. Therefore, we further propose to disentangle the adaptation procedure in the basic meta-learning training into two steps, the first one focusing on the depth sparsity while the second attending to the image content. During testing, we take the same strategy to adapt the model online to new multi-modal data. Experimental results and comprehensive ablations show that our MetaComp is capable of adapting to the depth completion in a new environment effectively and robust to changes in different modalities.
Establishing visual correspondence across images is a challenging and essential task. Recently, an influx of self-supervised methods have been proposed to better learn representations for visual correspondence. However, we find that these methods often fail to leverage semantic information and over-rely on the matching of low-level features. In contrast, human vision is capable of distinguishing between distinct objects as a pretext to tracking. Inspired by this paradigm, we propose to learn semantic-aware fine-grained correspondence. Firstly, we demonstrate that semantic correspondence is implicitly available through a rich set of image-level self-supervised methods. We further design a pixel-level self-supervised learning objective which specifically targets fine-grained correspondence. For downstream tasks, we fuse these two kinds of complementary correspondence representations together, demonstrating that they boost performance synergistically. Our method surpasses previous state-of-the-art self-supervised methods using convolutional networks on a variety of visual correspondence tasks, including video object segmentation, human pose tracking, and human part tracking.
In this paper, we examine self-supervised learning methods, particularly VICReg, to provide an information-theoretical understanding of their construction. As a first step, we demonstrate how information-theoretic quantities can be obtained for a deterministic network, offering a possible alternative to prior work that relies on stochastic models. This enables us to demonstrate how VICReg can be (re)discovered from first principles and its assumptions about data distribution. Furthermore, we empirically demonstrate the validity of our assumptions, confirming our novel understanding of VICReg. Finally, we believe that the derivation and insights we obtain can be generalized to many other SSL methods, opening new avenues for theoretical and practical understanding of SSL and transfer learning.
We introduce MABe22, a large-scale, multi-agent video and trajectory benchmark to assess the quality of learned behavior representations. This dataset is collected from a variety of biology experiments, and includes triplets of interacting mice (4.7 million frames video+pose tracking data, 10 million frames pose only), symbiotic beetle-ant interactions (10 million frames video data), and groups of interacting flies (4.4 million frames of pose tracking data). Accompanying these data, we introduce a panel of real-life downstream analysis tasks to assess the quality of learned representations by evaluating how well they preserve information about the experimental conditions (e.g. strain, time of day, optogenetic stimulation) and animal behavior. We test multiple state-of-the-art self-supervised video and trajectory representation learning methods to demonstrate the use of our benchmark, revealing that methods developed using human action datasets do not fully translate to animal datasets. We hope that our benchmark and dataset encourage a broader exploration of behavior representation learning methods across species and settings.
Automatic speech recognition (ASR) has progressed significantly in recent years due to the emergence of large-scale datasets and the self-supervised learning (SSL) paradigm. However, as its counterpart problem in the singing domain, the development of automatic lyric transcription (ALT) suffers from limited data and degraded intelligibility of sung lyrics. To fill in the performance gap between ALT and ASR, we attempt to exploit the similarities between speech and singing. In this work, we propose a transfer-learning-based ALT solution that takes advantage of these similarities by adapting wav2vec 2.0, an SSL ASR model, to the singing domain. We maximize the effectiveness of transfer learning by exploring the influence of different transfer starting points. We further enhance the performance by extending the original CTC model to a hybrid CTC/attention model. Our method surpasses previous approaches by a large margin on various ALT benchmark datasets. Further experiments show that, with even a tiny proportion of training data, our method still achieves competitive performance.
The fast development of self-supervised learning lowers the bar learning feature representation from massive unlabeled data and has triggered a series of research on change detection of remote sensing images. Challenges in adapting self-supervised learning from natural images classification to remote sensing images change detection arise from difference between the two tasks. The learned patch-level feature representations are not satisfying for the pixel-level precise change detection. In this paper, we proposed a novel pixel-level self-supervised hyperspectral spatial-spectral understanding network (HyperNet) to accomplish pixel-wise feature representation for effective hyperspectral change detection. Concretely, not patches but the whole images are fed into the network and the multi-temporal spatial-spectral features are compared pixel by pixel. Instead of processing the two-dimensional imaging space and spectral response dimension in hybrid style, a powerful spatial-spectral attention module is put forward to explore the spatial correlation and discriminative spectral features of multi-temporal hyperspectral images (HSIs), separately. Only the positive samples at the same location of bi-temporal HSIs are created and forced to be aligned, aiming at learning the spectral difference-invariant features. Moreover, a new similarity loss function named focal cosine is proposed to solve the problem of imbalanced easy and hard positive samples comparison, where the weights of those hard samples are enlarged and highlighted to promote the network training. Six hyperspectral datasets have been adopted to test the validity and generalization of proposed HyperNet. The extensive experiments demonstrate the superiority of HyperNet over the state-of-the-art algorithms on downstream hyperspectral change detection tasks.
With the improved performance of deep learning, the number of studies trying to apply deep learning to human emotion analysis is increasing rapidly. But even with this trend going on, it is still difficult to obtain high-quality images and annotations. For this reason, the Learning from Synthetic Data (LSD) Challenge, which learns from synthetic images and infers from real images, is one of the most interesting areas. In general, Domain Adaptation methods are widely used to address LSD challenges, but there is a limitation that target domains (real images) are still needed. Focusing on these limitations, we propose a framework Bootstrap Your Emotion Latent (BYEL), which uses only synthetic images in training. BYEL is implemented by adding Emotion Classifiers and Emotion Vector Subtraction to the BYOL framework that performs well in Self-Supervised Representation Learning. We train our framework using synthetic images generated from the Aff-wild2 dataset and evaluate it using real images from the Aff-wild2 dataset. The result shows that our framework (0.3084) performs 2.8% higher than the baseline (0.3) on the macro F1 score metric.
Feature selection is an important process in machine learning. It builds an interpretable and robust model by selecting the features that contribute the most to the prediction target. However, most mature feature selection algorithms, including supervised and semi-supervised, fail to fully exploit the complex potential structure between features. We believe that these structures are very important for the feature selection process, especially when labels are lacking and data is noisy. To this end, we innovatively introduce a deep learning-based self-supervised mechanism into feature selection problems, namely batch-Attention-based Self-supervision Feature Selection(A-SFS). Firstly, a multi-task self-supervised autoencoder is designed to uncover the hidden structure among features with the support of two pretext tasks. Guided by the integrated information from the multi-self-supervised learning model, a batch-attention mechanism is designed to generate feature weights according to batch-based feature selection patterns to alleviate the impacts introduced by a handful of noisy data. This method is compared to 14 major strong benchmarks, including LightGBM and XGBoost. Experimental results show that A-SFS achieves the highest accuracy in most datasets. Furthermore, this design significantly reduces the reliance on labels, with only 1/10 labeled data needed to achieve the same performance as those state of art baselines. Results show that A-SFS is also most robust to the noisy and missing data.