Self-supervised pre-training using so-called "pretext" tasks has recently shown impressive performance across a wide range of modalities. In this work, we advance self-supervised learning from permutations, by pre-training a model to reorder shuffled parts of the spectrogram of an audio signal, to improve downstream classification performance. We make two main contributions. First, we overcome the main challenges of integrating permutation inversions into an end-to-end training scheme, using recent advances in differentiable ranking. This was heretofore sidestepped by casting the reordering task as classification, fundamentally reducing the space of permutations that can be exploited. Our experiments validate that learning from all possible permutations improves the quality of the pre-trained representations over using a limited, fixed set. Second, we show that inverting permutations is a meaningful pretext task for learning audio representations in an unsupervised fashion. In particular, we improve instrument classification and pitch estimation of musical notes by reordering spectrogram patches in the time-frequency space.
The quality of the image representations obtained from self-supervised learning depends strongly on the type of data augmentations used in the learning formulation. Recent papers have ported these methods from still images to videos and found that leveraging both audio and video signals yields strong gains; however, they did not find that spatial augmentations such as cropping, which are very important for still images, work as well for videos. In this paper, we improve these formulations in two ways unique to the spatio-temporal aspect of videos. First, for space, we show that spatial augmentations such as cropping do work well for videos too, but that previous implementations, due to the high processing and memory cost, could not do this at a scale sufficient for it to work well. To address this issue, we first introduce Feature Crop, a method to simulate such augmentations much more efficiently directly in feature space. Second, we show that as opposed to naive average pooling, the use of transformer-based attention improves performance significantly, and is well suited for processing feature crops. Combining both of our discoveries into a new method, Space-Time Crop & Attend (STiCA) we achieve state-of-the-art performance across multiple video-representation learning benchmarks. In particular, we achieve new state-of-the-art accuracies of 67.0% on HMDB-51 and 93.1% on UCF-101 when pre-training on Kinetics-400.
While deep learning has enabled great advances in many areas of music, labeled music datasets remain especially hard, expensive, and time-consuming to create. In this work, we introduce SimCLR to the music domain and contribute a large chain of audio data augmentations to form a simple framework for self-supervised, contrastive learning of musical representations: CLMR. This approach works on raw time-domain music data and requires no labels to learn useful representations. We evaluate CLMR in the downstream task of music classification on the MagnaTagATune and Million Song datasets and present an ablation study to test which of our music-related innovations over SimCLR are most effective. A linear classifier trained on the proposed representations achieves a higher average precision than supervised models on the MagnaTagATune dataset, and performs comparably on the Million Song dataset. Moreover, we show that CLMR's representations are transferable using out-of-domain datasets, indicating that our method has strong generalisability in music classification. Lastly, we show that the proposed method allows data-efficient learning on smaller labeled datasets: we achieve an average precision of 33.1% despite using only 259 labeled songs in the MagnaTagATune dataset (1% of the full dataset) during linear evaluation. To foster reproducibility and future research on self-supervised learning in music, we publicly release the pre-trained models and the source code of all experiments of this paper.
Wav2vec-C introduces a novel representation learning technique combining elements from wav2vec 2.0 and VQ-VAE. Our model learns to reproduce quantized representations from partially masked speech encoding using a contrastive loss in a way similar to Wav2vec 2.0. However, the quantization process is regularized by an additional consistency network that learns to reconstruct the input features to the wav2vec 2.0 network from the quantized representations in a way similar to a VQ-VAE model. The proposed self-supervised model is trained on 10k hours of unlabeled data and subsequently used as the speech encoder in a RNN-T ASR model and fine-tuned with 1k hours of labeled data. This work is one of only a few studies of self-supervised learning on speech tasks with a large volume of real far-field labeled data. The Wav2vec-C encoded representations achieves, on average, twice the error reduction over baseline and a higher codebook utilization in comparison to wav2vec 2.0
Neural architecture search (NAS) has fostered various fields of machine learning. Despite its prominent dedications, many have criticized the intrinsic limitations of high computational cost. We aim to ameliorate this by proposing a pretraining scheme that can be generally applied to controller-based NAS. Our method, locality-based self-supervised classification task, leverages the structural similarity of network architectures to obtain good architecture representations. We incorporate our method into neural architecture optimization (NAO) to analyze the pretrained embeddings and its effectiveness and highlight that adding metric learning loss brings a favorable impact on NAS. Our code is available at \url{https://github.com/Multi-Objective-NAS/self-supervised-nas}.
Image clustering is a particularly challenging computer vision task, which aims to generate annotations without human supervision. Recent advances focus on the use of self-supervised learning strategies in image clustering, by first learning valuable semantics and then clustering the image representations. These multiple-phase algorithms, however, increase the computational time and their final performance is reliant on the first stage. By extending the self-supervised approach, we propose a novel single-phase clustering method that simultaneously learns meaningful representations and assigns the corresponding annotations. This is achieved by integrating a discrete representation into the self-supervised paradigm through a classifier net. Specifically, the proposed clustering objective employs mutual information, and maximizes the dependency between the integrated discrete representation and a discrete probability distribution. The discrete probability distribution is derived though the self-supervised process by comparing the learnt latent representation with a set of trainable prototypes. To enhance the learning performance of the classifier, we jointly apply the mutual information across multi-crop views. Our empirical results show that the proposed framework outperforms state-of-the-art techniques with the average accuracy of 89.1% and 49.0%, respectively, on CIFAR-10 and CIFAR-100/20 datasets. Finally, the proposed method also demonstrates attractive robustness to parameter settings, making it ready to be applicable to other datasets.
In this paper, we present self-supervised shared latent embedding (S3LE), a data-driven motion retargeting method that enables the generation of natural motions in humanoid robots from motion capture data or RGB videos. While it requires paired data consisting of human poses and their corresponding robot configurations, it significantly alleviates the necessity of time-consuming data-collection via novel paired data generating processes. Our self-supervised learning procedure consists of two steps: automatically generating paired data to bootstrap the motion retargeting, and learning a projection-invariant mapping to handle the different expressivity of humans and humanoid robots. Furthermore, our method guarantees that the generated robot pose is collision-free and satisfies position limits by utilizing nonparametric regression in the shared latent space. We demonstrate that our method can generate expressive robotic motions from both the CMU motion capture database and YouTube videos.
Since we can leverage a large amount of unlabeled data without any human supervision to train a model and transfer the knowledge to target tasks, self-supervised learning is a de-facto component for the recent success of deep learning in various fields. However, in many cases, there is a discrepancy between a self-supervised learning objective and a task-specific objective. In order to tackle such discrepancy in Text-to-SQL task, we propose a novel self-supervised learning framework. We utilize the task-specific properties of Text-to-SQL task and the underlying structures of table contents to train the models to learn useful knowledge of the \textit{header-column} alignment task from unlabeled table data. We are able to transfer the knowledge to the supervised Text-to-SQL training with annotated samples, so that the model can leverage the knowledge to better perform the \textit{header-span} alignment task to predict SQL statements. Experimental results show that our self-supervised learning framework significantly improves the performance of the existing strong BERT based models without using large external corpora. In particular, our method is effective for training the model with scarce labeled data. The source code of this work is available in GitHub.
Meta-reinforcement learning typically requires orders of magnitude more samples than single task reinforcement learning methods. This is because meta-training needs to deal with more diverse distributions and train extra components such as context encoders. To address this, we propose a novel self-supervised learning task, which we named Trajectory Contrastive Learning (TCL), to improve meta-training. TCL adopts contrastive learning and trains a context encoder to predict whether two transition windows are sampled from the same trajectory. TCL leverages the natural hierarchical structure of context-based meta-RL and makes minimal assumptions, allowing it to be generally applicable to context-based meta-RL algorithms. It accelerates the training of context encoders and improves meta-training overall. Experiments show that TCL performs better or comparably than a strong meta-RL baseline in most of the environments on both meta-RL MuJoCo (5 of 6) and Meta-World benchmarks (44 out of 50).
Inspired by the recent progress in self-supervised learning for computer vision that generates supervision using data augmentations, we explore a new general-purpose audio representation learning approach. We propose learning general-purpose audio representation from a single audio segment without expecting relationships between different time segments of audio samples. To implement this principle, we introduce Bootstrap Your Own Latent (BYOL) for Audio (BYOL-A, pronounced "viola"), an audio self-supervised learning method based on BYOL for learning general-purpose audio representation. Unlike most previous audio self-supervised learning methods that rely on agreement of vicinity audio segments or disagreement of remote ones, BYOL-A creates contrasts in an augmented audio segment pair derived from a single audio segment. With a combination of normalization and augmentation techniques, BYOL-A achieves state-of-the-art results in various downstream tasks. Extensive ablation studies also clarified the contribution of each component and their combinations.
In this paper, we study the cross-modal image retrieval, where the inputs contain a source image plus some text that describes certain modifications to this image and the desired image. Prior work usually uses a three-stage strategy to tackle this task: 1) extract the features of the inputs; 2) fuse the feature of the source image and its modified text to obtain fusion feature; 3) learn a similarity metric between the desired image and the source image + modified text by using deep metric learning. Since classical image/text encoders can learn the useful representation and common pair-based loss functions of distance metric learning are enough for cross-modal retrieval, people usually improve retrieval accuracy by designing new fusion networks. However, these methods do not successfully handle the modality gap caused by the inconsistent distribution and representation of the features of different modalities, which greatly influences the feature fusion and similarity learning. To alleviate this problem, we adopt the contrastive self-supervised learning method Deep InforMax (DIM) to our approach to bridge this gap by enhancing the dependence between the text, the image, and their fusion. Specifically, our method narrows the modality gap between the text modality and the image modality by maximizing mutual information between their not exactly semantically identical representation. Moreover, we seek an effective common subspace for the semantically same fusion feature and desired image's feature by utilizing Deep InforMax between the low-level layer of the image encoder and the high-level layer of the fusion network. Extensive experiments on three large-scale benchmark datasets show that we have bridged the modality gap between different modalities and achieve state-of-the-art retrieval performance.
Few-shot learning is an interesting and challenging study, which enables machines to learn from few samples like humans. Existing studies rarely exploit auxiliary information from large amount of unlabeled data. Self-supervised learning is emerged as an efficient method to utilize unlabeled data. Existing self-supervised learning methods always rely on the combination of geometric transformations for the single sample by augmentation, while seriously neglect the endogenous correlation information among different samples that is the same important for the task. In this work, we propose a Graph-driven Clustering (GC), a novel augmentation-free method for self-supervised learning, which does not rely on any auxiliary sample and utilizes the endogenous correlation information among input samples. Besides, we propose Multi-pretext Attention Network (MAN), which exploits a specific attention mechanism to combine the traditional augmentation-relied methods and our GC, adaptively learning their optimized weights to improve the performance and enabling the feature extractor to obtain more universal representations. We evaluate our MAN extensively on miniImageNet and tieredImageNet datasets and the results demonstrate that the proposed method outperforms the state-of-the-art (SOTA) relevant methods.
Self-supervised learning has been widely used to obtain transferrable representations from unlabeled images. Especially, recent contrastive learning methods have shown impressive performances on downstream image classification tasks. While these contrastive methods mainly focus on generating invariant global representations at the image-level under semantic-preserving transformations, they are prone to overlook spatial consistency of local representations and therefore have a limitation in pretraining for localization tasks such as object detection and instance segmentation. Moreover, aggressively cropped views used in existing contrastive methods can minimize representation distances between the semantically different regions of a single image. In this paper, we propose a spatially consistent representation learning algorithm (SCRL) for multi-object and location-specific tasks. In particular, we devise a novel self-supervised objective that tries to produce coherent spatial representations of a randomly cropped local region according to geometric translations and zooming operations. On various downstream localization tasks with benchmark datasets, the proposed SCRL shows significant performance improvements over the image-level supervised pretraining as well as the state-of-the-art self-supervised learning methods. Code is available at https://github.com/kakaobrain/scrl
Despite their irresistible success, deep learning algorithms still heavily rely on annotated data. On the other hand, unsupervised settings pose many challenges, especially about determining the right inductive bias in diverse scenarios. One scalable solution is to make the model generate the supervision for itself by leveraging some part of the input data, which is known as self-supervised learning. In this paper, we use the prediction task as self-supervision and build a novel object-centric model for image sequence representation. In addition to disentangling the notion of objects and the motion dynamics, our compositional structure explicitly handles occlusion and inpaints inferred objects and background for the composition of the predicted frame. With the aid of auxiliary loss functions that promote spatially and temporally consistent object representations, our self-supervised framework can be trained without the help of any manual annotation or pretrained network. Initial experiments confirm that the proposed pipeline is a promising step towards object-centric video prediction.
Text classification is a widely studied problem and has broad applications. In many real-world problems, the number of texts for training classification models is limited, which renders these models prone to overfitting. To address this problem, we propose SSL-Reg, a data-dependent regularization approach based on self-supervised learning (SSL). SSL is an unsupervised learning approach which defines auxiliary tasks on input data without using any human-provided labels and learns data representations by solving these auxiliary tasks. In SSL-Reg, a supervised classification task and an unsupervised SSL task are performed simultaneously. The SSL task is unsupervised, which is defined purely on input texts without using any human-provided labels. Training a model using an SSL task can prevent the model from being overfitted to a limited number of class labels in the classification task. Experiments on 17 text classification datasets demonstrate the effectiveness of our proposed method.
Most change detection methods assume that pre-change and post-change images are acquired by the same sensor. However, in many real-life scenarios, e.g., natural disaster, it is more practical to use the latest available images before and after the occurrence of incidence, which may be acquired using different sensors. In particular, we are interested in the combination of the images acquired by optical and Synthetic Aperture Radar (SAR) sensors. SAR images appear vastly different from the optical images even when capturing the same scene. Adding to this, change detection methods are often constrained to use only target image-pair, no labeled data, and no additional unlabeled data. Such constraints limit the scope of traditional supervised machine learning and unsupervised generative approaches for multi-sensor change detection. Recent rapid development of self-supervised learning methods has shown that some of them can even work with only few images. Motivated by this, in this work we propose a method for multi-sensor change detection using only the unlabeled target bi-temporal images that are used for training a network in self-supervised fashion by using deep clustering and contrastive learning. The proposed method is evaluated on four multi-modal bi-temporal scenes showing change and the benefits of our self-supervised approach are demonstrated.
Contrastive learning is a key technique of modern self-supervised learning. The broader accessibility of earlier approaches is hindered by the need of heavy computational resources (e.g., at least 8 GPUs or 32 TPU cores), which accommodate for large-scale negative samples or momentum. The more recent SimSiam approach addresses such key limitations via stop-gradient without momentum encoders. In medical image analysis, multiple instances can be achieved from the same patient or tissue. Inspired by these advances, we propose a simple triplet representation learning (SimTriplet) approach on pathological images. The contribution of the paper is three-fold: (1) The proposed SimTriplet method takes advantage of the multi-view nature of medical images beyond self-augmentation; (2) The method maximizes both intra-sample and inter-sample similarities via triplets from positive pairs, without using negative samples; and (3) The recent mix precision training is employed to advance the training by only using a single GPU with 16GB memory. By learning from 79,000 unlabeled pathological patch images, SimTriplet achieved 10.58% better performance compared with supervised learning. It also achieved 2.13% better performance compared with SimSiam. Our proposed SimTriplet can achieve decent performance using only 1% labeled data. The code and data are available at https://github.com/hrlblab/SimTriple.
The effective combination of the complementary information provided by the huge amount of unlabeled multi-sensor data (e.g., Synthetic Aperture Radar (SAR) and optical images) is a critical topic in remote sensing. Recently, contrastive learning methods have reached remarkable success in obtaining meaningful feature representations from multi-view data. However, these methods only focus on image-level features, which may not satisfy the requirement for dense prediction tasks such as land-cover mapping. In this work, we propose a self-supervised framework for SAR-optical data fusion and land-cover mapping tasks. SAR and optical images are fused by using multi-view contrastive loss at image-level and super-pixel level in the early, intermediate and later fashion individually. For the land-cover mapping task, we assign each pixel a land-cover class by the joint use of pre-trained features and spectral information of the image itself. Experimental results show that the proposed approach achieves a comparable accuracy and that reduces the dimension of features with respect to the image-level contrastive learning method. Among three fusion fashions, the intermediate fusion strategy achieves the best performance. The combination of the pixel-level fusion approach and spectral indices leads to further improvements on the land-cover mapping task with respect to the image-level fusion approach, especially with few pseudo labels.
Computing power, big data, and advancement of algorithms have led to a renewed interest in artificial intelligence (AI), especially in deep learning (DL). The success of DL largely lies on data representation because different representations can indicate to a degree the different explanatory factors of variation behind the data. In the last few year, the most successful story in DL is supervised learning. However, to apply supervised learning, one challenge is that data labels are expensive to get, noisy, or only partially available. With consideration that we human beings learn in an unsupervised way; self-supervised learning methods have garnered a lot of attention recently. A dominant force in self-supervised learning is the autoencoder, which has multiple uses (e.g., data representation, anomaly detection, denoise). This research explored the application of an autoencoder to learn effective data representation of helicopter flight track data, and then to support helicopter track identification. Our testing results are promising. For example, at Phoenix Deer Valley (DVT) airport, where 70% of recorded flight tracks have missing aircraft types, the autoencoder can help to identify twenty-two times more helicopters than otherwise detectable using rule-based methods; for Grand Canyon West Airport (1G4) airport, the autoencoder can identify thirteen times more helicopters than a current rule-based approach. Our approach can also identify mislabeled aircraft types in the flight track data and find true types for records with pseudo aircraft type labels such as HELO. With improved labelling, studies using these data sets can produce more reliable results.
Radiomic representations can quantify properties of regions of interest in medical image data. Classically, they account for pre-defined statistics of shape, texture, and other low-level image features. Alternatively, deep learning-based representations are derived from supervised learning but require expensive annotations from experts and often suffer from overfitting and data imbalance issues. In this work, we address the challenge of learning representations of 3D medical images for an effective quantification under data imbalance. We propose a \emph{self-supervised} representation learning framework to learn high-level features of 3D volumes as a complement to existing radiomics features. Specifically, we demonstrate how to learn image representations in a self-supervised fashion using a 3D Siamese network. More importantly, we deal with data imbalance by exploiting two unsupervised strategies: a) sample re-weighting, and b) balancing the composition of training batches. When combining our learned self-supervised feature with traditional radiomics, we show significant improvement in brain tumor classification and lung cancer staging tasks covering MRI and CT imaging modalities.