The success of deep learning comes from its ability to capture the hierarchical structure of data by learning high-level representations defined in terms of low-level ones. In this paper we explore self-supervised learning of hierarchical representations of speech by applying multiple levels of Contrastive Predictive Coding (CPC). We observe that simply stacking two CPC models does not yield significant improvements over single-level architectures. Inspired by the fact that speech is often described as a sequence of discrete units unevenly distributed in time, we propose a model in which the output of a low-level CPC module is non-uniformly downsampled to directly minimize the loss of a high-level CPC module. The latter is designed to also enforce a prior of separability and discreteness in its representations by enforcing dissimilarity of successive high-level representations through focused negative sampling, and by quantization of the prediction targets. Accounting for the structure of the speech signal improves upon single-level CPC features and enhances the disentanglement of the learned representations, as measured by downstream speech recognition tasks, while resulting in a meaningful segmentation of the signal that closely resembles phone boundaries.
We address the problem of evaluating the quality of self-supervised learning (SSL) models without access to supervised labels, while being agnostic to the architecture, learning algorithm or data manipulation used during training. We argue that representations can be evaluated through the lens of expressiveness and learnability. We propose to use the Intrinsic Dimension (ID) to assess expressiveness and introduce Cluster Learnability (CL) to assess learnability. CL is measured in terms of the performance of a KNN classifier trained to predict labels obtained by clustering the representations with K-means. We thus combine CL and ID into a single predictor -- CLID. Through a large-scale empirical study with a diverse family of SSL algorithms, we find that CLID better correlates with in-distribution model performance than other competing recent evaluation schemes. We also benchmark CLID on out-of-domain generalization, where CLID serves as a predictor of the transfer performance of SSL models on several visual classification tasks, yielding improvements with respect to the competing baselines.
Several deep neural networks have recently been shown to generate activations similar to those of the brain in response to the same input. These algorithms, however, remain largely implausible: they require (1) extraordinarily large amounts of data, (2) unobtainable supervised labels, (3) textual rather than raw sensory input, and / or (4) implausibly large memory (e.g. thousands of contextual words). These elements highlight the need to identify algorithms that, under these limitations, would suffice to account for both behavioral and brain responses. Focusing on the issue of speech processing, we here hypothesize that self-supervised algorithms trained on the raw waveform constitute a promising candidate. Specifically, we compare a recent self-supervised architecture, Wav2Vec 2.0, to the brain activity of 412 English, French, and Mandarin individuals recorded with functional Magnetic Resonance Imaging (fMRI), while they listened to ~1h of audio books. Our results are four-fold. First, we show that this algorithm learns brain-like representations with as little as 600 hours of unlabelled speech -- a quantity comparable to what infants can be exposed to during language acquisition. Second, its functional hierarchy aligns with the cortical hierarchy of speech processing. Third, different training regimes reveal a functional specialization akin to the cortex: Wav2Vec 2.0 learns sound-generic, speech-specific and language-specific representations similar to those of the prefrontal and temporal cortices. Fourth, we confirm the similarity of this specialization with the behavior of 386 additional participants. These elements, resulting from the largest neuroimaging benchmark to date, show how self-supervised learning can account for a rich organization of speech processing in the brain, and thus delineate a path to identify the laws of language acquisition which shape the human brain.
Graph contrastive learning (GCL) alleviates the heavy reliance on label information for graph representation learning (GRL) via self-supervised learning schemes. The core idea is to learn by maximising mutual information for similar instances, which requires similarity computation between two node instances. However, GCL is inefficient in both time and memory consumption. In addition, GCL normally requires a large number of training epochs to be well-trained on large-scale datasets. Inspired by an observation of a technical defect (i.e., inappropriate usage of Sigmoid function) commonly used in two representative GCL works, DGI and MVGRL, we revisit GCL and introduce a new learning paradigm for self-supervised graph representation learning, namely, Group Discrimination (GD), and propose a novel GD-based method called Graph Group Discrimination (GGD). Instead of similarity computation, GGD directly discriminates two groups of node samples with a very simple binary cross-entropy loss. In addition, GGD requires much fewer training epochs to obtain competitive performance compared with GCL methods on large-scale datasets. These two advantages endow GGD with very efficient property. Extensive experiments show that GGD outperforms state-of-the-art self-supervised methods on eight datasets. In particular, GGD can be trained in 0.18 seconds (6.44 seconds including data preprocessing) on ogbn-arxiv, which is orders of magnitude (10,000+) faster than GCL baselines while consuming much less memory. Trained with 9 hours on ogbn-papers100M with billion edges, GGD outperforms its GCL counterparts in both accuracy and efficiency.
While the empirical success of self-supervised learning (SSL) heavily relies on the usage of deep nonlinear models, existing theoretical works on SSL understanding still focus on linear ones. In this paper, we study the role of nonlinearity in the training dynamics of contrastive learning (CL) on one and two-layer nonlinear networks with homogeneous activation $h(x) = h'(x)x$. We have two major theoretical discoveries. First, the presence of nonlinearity can lead to many local optima even in 1-layer setting, each corresponding to certain patterns from the data distribution, while with linear activation, only one major pattern can be learned. This suggests that models with lots of parameters can be regarded as a \emph{brute-force} way to find these local optima induced by nonlinearity. Second, in the 2-layer case, linear activation is proven not capable of learning specialized weights into diverse patterns, demonstrating the importance of nonlinearity. In addition, for 2-layer setting, we also discover \emph{global modulation}: those local patterns discriminative from the perspective of global-level patterns are prioritized to learn, further characterizing the learning process. Simulation verifies our theoretical findings.
Learning from audio-visual data offers many possibilities to express correspondence between the audio and visual content, similar to the human perception that relates aural and visual information. In this work, we present a method for self-supervised representation learning based on audio-visual spatial alignment (AVSA), a more sophisticated alignment task than the audio-visual correspondence (AVC). In addition to the correspondence, AVSA also learns from the spatial location of acoustic and visual content. Based on 360$^\text{o}$ video and Ambisonics audio, we propose selection of visual objects using object detection, and beamforming of the audio signal towards the detected objects, attempting to learn the spatial alignment between objects and the sound they produce. We investigate the use of spatial audio features to represent the audio input, and different audio formats: Ambisonics, mono, and stereo. Experimental results show a 10 $\%$ improvement on AVSA for the first order ambisonics intensity vector (FOA-IV) in comparison with log-mel spectrogram features; the addition of object-oriented crops also brings significant performance increases for the human action recognition downstream task. A number of audio-only downstream tasks are devised for testing the effectiveness of the learnt audio feature representation, obtaining performance comparable to state-of-the-art methods on acoustic scene classification from ambisonic and binaural audio.
Self-supervised learning (SSL) has delivered superior performance on a variety of downstream vision tasks. Two main-stream SSL frameworks have been proposed, i.e., Instance Discrimination (ID) and Masked Image Modeling (MIM). ID pulls together representations from different views of the same image, while avoiding feature collapse. It lacks spatial sensitivity, which requires modeling the local structure within each image. On the other hand, MIM reconstructs the original content given a masked image. It instead does not have good semantic alignment, which requires projecting semantically similar views into nearby representations. To address this dilemma, we observe that (1) semantic alignment can be achieved by matching different image views with strong augmentations; (2) spatial sensitivity can benefit from predicting dense representations with masked images. Driven by these analysis, we propose Siamese Image Modeling (SiameseIM), which predicts the dense representations of an augmented view, based on another masked view from the same image but with different augmentations. SiameseIM uses a Siamese network with two branches. The online branch encodes the first view, and predicts the second view's representation according to the relative positions between these two views. The target branch produces the target by encoding the second view. SiameseIM can surpass both ID and MIM on a wide range of downstream tasks, including ImageNet finetuning and linear probing, COCO and LVIS detection, and ADE20k semantic segmentation. The improvement is more significant in few-shot, long-tail and robustness-concerned scenarios. Code shall be released at https://github.com/fundamentalvision/Siamese-Image-Modeling.
Self-supervised learning for computer vision has achieved tremendous progress and improved many downstream vision tasks such as image classification, semantic segmentation, and object detection. Among these, generative self-supervised vision learning approaches such as MAE and BEiT show promising performance. However, their global masked reconstruction mechanism is computationally demanding. To address this issue, we propose local masked reconstruction (LoMaR), a simple yet effective approach that performs masked reconstruction within a small window of 7$\times$7 patches on a simple Transformer encoder, improving the trade-off between efficiency and accuracy compared to global masked reconstruction over the entire image. Extensive experiments show that LoMaR reaches 84.1% top-1 accuracy on ImageNet-1K classification, outperforming MAE by 0.5%. After finetuning the pretrained LoMaR on 384$\times$384 images, it can reach 85.4% top-1 accuracy, surpassing MAE by 0.6%. On MS COCO, LoMaR outperforms MAE by 0.5 $\text{AP}^\text{box}$ on object detection and 0.5 $\text{AP}^\text{mask}$ on instance segmentation. LoMaR is especially more computation-efficient on pretraining high-resolution images, e.g., it is 3.1$\times$ faster than MAE with 0.2% higher classification accuracy on pretraining 448$\times$448 images. This local masked reconstruction learning mechanism can be easily integrated into any other generative self-supervised learning approach. Our code is publicly available in https://github.com/junchen14/LoMaR.
Recent advances in contrastive learning have enlightened diverse applications across various semi-supervised fields. Jointly training supervised learning and unsupervised learning with a shared feature encoder becomes a common scheme. Though it benefits from taking advantage of both feature-dependent information from self-supervised learning and label-dependent information from supervised learning, this scheme remains suffering from bias of the classifier. In this work, we systematically explore the relationship between self-supervised learning and supervised learning, and study how self-supervised learning helps robust data-efficient deep learning. We propose hyperspherical consistency regularization (HCR), a simple yet effective plug-and-play method, to regularize the classifier using feature-dependent information and thus avoid bias from labels. Specifically, HCR first projects logits from the classifier and feature projections from the projection head on the respective hypersphere, then it enforces data points on hyperspheres to have similar structures by minimizing binary cross entropy of pairwise distances' similarity metrics. Extensive experiments on semi-supervised and weakly-supervised learning demonstrate the effectiveness of our method, by showing superior performance with HCR.
With the recent promising results of contrastive learning in the self-supervised learning paradigm, supervised contrastive learning has successfully extended these contrastive approaches to supervised contexts, outperforming cross-entropy on various datasets. However, supervised contrastive learning inherently employs label information in a binary form--either positive or negative--using a one-hot target vector. This structure struggles to adapt to methods that exploit label information as a probability distribution, such as CutMix and knowledge distillation. In this paper, we introduce a generalized supervised contrastive loss, which measures cross-entropy between label similarity and latent similarity. This concept enhances the capabilities of supervised contrastive loss by fully utilizing the label distribution and enabling the adaptation of various existing techniques for training modern neural networks. Leveraging this generalized supervised contrastive loss, we construct a tailored framework: the Generalized Supervised Contrastive Learning (GenSCL). Compared to existing contrastive learning frameworks, GenSCL incorporates additional enhancements, including advanced image-based regularization techniques and an arbitrary teacher classifier. When applied to ResNet50 with the Momentum Contrast technique, GenSCL achieves a top-1 accuracy of 77.3% on ImageNet, a 4.1% relative improvement over traditional supervised contrastive learning. Moreover, our method establishes new state-of-the-art accuracies of 98.2% and 87.0% on CIFAR10 and CIFAR100 respectively when applied to ResNet50, marking the highest reported figures for this architecture.
One of the largest problems in medical image processing is the lack of annotated data. Labeling medical images often requires highly trained experts and can be a time-consuming process. In this paper, we evaluate a method of reducing the need for labeled data in medical image object detection by using self-supervised neural network pretraining. We use a dataset of chest X-ray images with bounding box labels for 13 different classes of anomalies. The networks are pretrained on a percentage of the dataset without labels and then fine-tuned on the rest of the dataset. We show that it is possible to achieve similar performance to a fully supervised model in terms of mean average precision and accuracy with only 60\% of the labeled data. We also show that it is possible to increase the maximum performance of a fully-supervised model by adding a self-supervised pretraining step, and this effect can be observed with even a small amount of unlabeled data for pretraining.
The computational drug repositioning aims to discover new uses for marketed drugs, which can accelerate the drug development process and play an important role in the existing drug discovery system. However, the number of validated drug-disease associations is scarce compared to the number of drugs and diseases in the real world. Too few labeled samples will make the classification model unable to learn effective latent factors of drugs, resulting in poor generalization performance. In this work, we propose a multi-task self-supervised learning framework for computational drug repositioning. The framework tackles label sparsity by learning a better drug representation. Specifically, we take the drug-disease association prediction problem as the main task, and the auxiliary task is to use data augmentation strategies and contrast learning to mine the internal relationships of the original drug features, so as to automatically learn a better drug representation without supervised labels. And through joint training, it is ensured that the auxiliary task can improve the prediction accuracy of the main task. More precisely, the auxiliary task improves drug representation and serving as additional regularization to improve generalization. Furthermore, we design a multi-input decoding network to improve the reconstruction ability of the autoencoder model. We evaluate our model using three real-world datasets. The experimental results demonstrate the effectiveness of the multi-task self-supervised learning framework, and its predictive ability is superior to the state-of-the-art model.
Vision Transformers (ViTs) enabled the use of the transformer architecture on vision tasks showing impressive performances when trained on big datasets. However, on relatively small datasets, ViTs are less accurate given their lack of inductive bias. To this end, we propose a simple but still effective Self-Supervised Learning (SSL) strategy to train ViTs, that without any external annotation or external data, can significantly improve the results. Specifically, we define a set of SSL tasks based on relations of image patches that the model has to solve before or jointly the supervised task. Differently from ViT, our RelViT model optimizes all the output tokens of the transformer encoder that are related to the image patches, thus exploiting more training signals at each training step. We investigated our methods on several image benchmarks finding that RelViT improves the SSL state-of-the-art methods by a large margin, especially on small datasets. Code is available at: https://github.com/guglielmocamporese/relvit.
Self-supervised learning aims to learn a embedding space where semantically similar samples are close. Contrastive learning methods pull views of samples together and push different samples away, which utilizes semantic invariance of augmentation but ignores the relationship between samples. To better exploit the power of augmentation, we observe that semantically similar samples are more likely to have similar augmented views. Therefore, we can take the augmented views as a special description of a sample. In this paper, we model such a description as the augmentation distribution and we call it augmentation feature. The similarity in augmentation feature reflects how much the views of two samples overlap and is related to their semantical similarity. Without computational burdens to explicitly estimate values of the augmentation feature, we propose Augmentation Component Analysis (ACA) with a contrastive-like loss to learn principal components and an on-the-fly projection loss to embed data. ACA equals an efficient dimension reduction by PCA and extracts low-dimensional embeddings, theoretically preserving the similarity of augmentation distribution between samples. Empirical results show our method can achieve competitive results against various traditional contrastive learning methods on different benchmarks.
Bipartite graphs are powerful data structures to model interactions between two types of nodes, which have been used in a variety of applications, such as recommender systems, information retrieval, and drug discovery. A fundamental challenge for bipartite graphs is how to learn informative node embeddings. Despite the success of recent self-supervised learning methods on bipartite graphs, their objectives are discriminating instance-wise positive and negative node pairs, which could contain cluster-level errors. In this paper, we introduce a novel co-cluster infomax (COIN) framework, which captures the cluster-level information by maximizing the mutual information of co-clusters. Different from previous infomax methods which estimate mutual information by neural networks, COIN could easily calculate mutual information. Besides, COIN is an end-to-end coclustering method which can be trained jointly with other objective functions and optimized via back-propagation. Furthermore, we also provide theoretical analysis for COIN. We theoretically prove that COIN is able to effectively increase the mutual information of node embeddings and COIN is upper-bounded by the prior distributions of nodes. We extensively evaluate the proposed COIN framework on various benchmark datasets and tasks to demonstrate the effectiveness of COIN.
Computer-aided diagnosis (CAD) can help pathologists improve diagnostic accuracy together with consistency and repeatability for cancers. However, the CAD models trained with the histopathological images only from a single center (hospital) generally suffer from the generalization problem due to the straining inconsistencies among different centers. In this work, we propose a pseudo-data based self-supervised federated learning (FL) framework, named SSL-FT-BT, to improve both the diagnostic accuracy and generalization of CAD models. Specifically, the pseudo histopathological images are generated from each center, which contains inherent and specific properties corresponding to the real images in this center, but does not include the privacy information. These pseudo images are then shared in the central server for self-supervised learning (SSL). A multi-task SSL is then designed to fully learn both the center-specific information and common inherent representation according to the data characteristics. Moreover, a novel Barlow Twins based FL (FL-BT) algorithm is proposed to improve the local training for the CAD model in each center by conducting contrastive learning, which benefits the optimization of the global model in the FL procedure. The experimental results on three public histopathological image datasets indicate the effectiveness of the proposed SSL-FL-BT on both diagnostic accuracy and generalization.
Text-to-Speech (TTS) has recently seen great progress in synthesizing high-quality speech owing to the rapid development of parallel TTS systems, but producing speech with naturalistic prosodic variations, speaking styles and emotional tones remains challenging. Moreover, since duration and speech are generated separately, parallel TTS models still have problems finding the best monotonic alignments that are crucial for naturalistic speech synthesis. Here, we propose StyleTTS, a style-based generative model for parallel TTS that can synthesize diverse speech with natural prosody from a reference speech utterance. With novel Transferable Monotonic Aligner (TMA) and duration-invariant data augmentation schemes, our method significantly outperforms state-of-the-art models on both single and multi-speaker datasets in subjective tests of speech naturalness and speaker similarity. Through self-supervised learning of the speaking styles, our model can synthesize speech with the same prosodic and emotional tone as any given reference speech without the need for explicitly labeling these categories.
In the field of post-disaster assessment, for timely and accurate rescue and localization after a disaster, people need to know the location of damaged buildings. In deep learning, some scholars have proposed methods to make automatic and highly accurate building damage assessments by remote sensing images, which are proved to be more efficient than assessment by domain experts. However, due to the lack of a large amount of labeled data, these kinds of tasks can suffer from being able to do an accurate assessment, as the efficiency of deep learning models relies highly on labeled data. Although existing semi-supervised and unsupervised studies have made breakthroughs in this area, none of them has completely solved this problem. Therefore, we propose adopting a self-supervised comparative learning approach to address the task without the requirement of labeled data. We constructed a novel asymmetric twin network architecture and tested its performance on the xBD dataset. Experiment results of our model show the improvement compared to baseline and commonly used methods. We also demonstrated the potential of self-supervised methods for building damage recognition awareness.
Self-supervised learning on large-scale Vision Transformers (ViTs) as pre-training methods has achieved promising downstream performance. Yet, how much these pre-training paradigms promote lightweight ViTs' performance is considerably less studied. In this work, we develop and benchmark several self-supervised pre-training methods on image classification tasks and some downstream dense prediction tasks. We surprisingly find that if proper pre-training is adopted, even vanilla lightweight ViTs show comparable performance to previous SOTA networks with delicate architecture design. It breaks the recently popular conception that vanilla ViTs are not suitable for vision tasks in lightweight regimes. We also point out some defects of such pre-training, e.g., failing to benefit from large-scale pre-training data and showing inferior performance on data-insufficient downstream tasks. Furthermore, we analyze and clearly show the effect of such pre-training by analyzing the properties of the layer representation and attention maps for related models. Finally, based on the above analyses, a distillation strategy during pre-training is developed, which leads to further downstream performance improvement for MAE-based pre-training. Code is available at https://github.com/wangsr126/mae-lite.
Supervised deep neural networks are the-state-of-the-art for many tasks in the remote sensing domain, against the fact that such techniques require the dataset consisting of pairs of input and label, which are rare and expensive to collect in term of both manpower and resources. On the other hand, there are abundance of raw satellite images available both for commercial and academic purposes. Hence, in this work, we tackle the insufficient labeled data problem in satellite image classification task by introducing the process based on the self-supervised learning technique to create the synthetic labels for satellite image patches. These synthetic labels can be used as the training dataset for the existing supervised learning techniques. In our experiments, we show that the models trained on the synthetic labels give similar performance to the models trained on the real labels. And in the process of creating the synthetic labels, we also obtain the visual representation vectors that are versatile and knowledge transferable.