Video anomaly detection is an ill-posed problem because it relies on many parameters such as appearance, pose, camera angle, background, and more. We distill the problem to anomaly detection of human pose, thus decreasing the risk of nuisance parameters such as appearance affecting the result. Focusing on pose alone also has the side benefit of reducing bias against distinct minority groups. Our model works directly on human pose graph sequences and is exceptionally lightweight (~1K parameters), capable of running on any machine able to run the pose estimation with negligible additional resources. We leverage the highly compact pose representation in a normalizing flows framework, which we extend to tackle the unique characteristics of spatio-temporal pose data and show its advantages in this use case. The algorithm is quite general and can handle training data of only normal examples as well as a supervised setting that consists of labeled normal and abnormal examples. We report state-of-the-art results on two anomaly detection benchmarks - the unsupervised ShanghaiTech dataset and the recent supervised UBnormal dataset.
Despite impressive performance for high-level downstream tasks, self-supervised pre-training methods have not yet fully delivered on dense geometric vision tasks such as stereo matching or optical flow. The application of self-supervised concepts, such as instance discrimination or masked image modeling, to geometric tasks is an active area of research. In this work, we build on the recent cross-view completion framework, a variation of masked image modeling that leverages a second view from the same scene which makes it well suited for binocular downstream tasks. The applicability of this concept has so far been limited in at least two ways: (a) by the difficulty of collecting real-world image pairs -- in practice only synthetic data have been used -- and (b) by the lack of generalization of vanilla transformers to dense downstream tasks for which relative position is more meaningful than absolute position. We explore three avenues of improvement. First, we introduce a method to collect suitable real-world image pairs at large scale. Second, we experiment with relative positional embeddings and show that they enable vision transformers to perform substantially better. Third, we scale up vision transformer based cross-completion architectures, which is made possible by the use of large amounts of data. With these improvements, we show for the first time that state-of-the-art results on stereo matching and optical flow can be reached without using any classical task-specific techniques like correlation volume, iterative estimation, image warping or multi-scale reasoning, thus paving the way towards universal vision models.
Transformer-based self-supervised models have achieved remarkable success in speech processing, but their large size and high inference cost present significant challenges for real-world deployment. While numerous compression techniques have been proposed, inconsistent evaluation metrics make it difficult to compare their practical effectiveness. In this work, we conduct a comprehensive study of four common compression methods, including weight pruning, head pruning, low-rank approximation, and knowledge distillation on self-supervised speech Transformers. We evaluate each method under three key metrics: parameter count, multiply-accumulate operations, and real-time factor. Results show that each method offers distinct advantages. In addition, we contextualize recent compression techniques, comparing DistilHuBERT, FitHuBERT, LightHuBERT, ARMHuBERT, and STaRHuBERT under the same framework, offering practical guidance on compression for deployment.
Self-supervised models have had great success in learning speech representations that can generalize to various downstream tasks. However, most self-supervised models require a large amount of compute and multiple GPUs to train, significantly hampering the development of self-supervised learning. In an attempt to reduce the computation of training, we revisit the training of HuBERT, a highly successful self-supervised model. We improve and simplify several key components, including the loss function, input representation, and training in multiple stages. Our model, MelHuBERT, is able to achieve favorable performance on phone recognition, speaker identification, and automatic speech recognition against HuBERT, while saving 31.2% of the pre-training time, or equivalently 33.5% MACs per one second speech. The code and pre-trained models are available in https://github.com/nervjack2/MelHuBERT.
The superior performance of modern deep networks usually comes with a costly training procedure. This paper presents a new curriculum learning approach for the efficient training of visual backbones (e.g., vision Transformers). Our work is inspired by the inherent learning dynamics of deep networks: we experimentally show that at an earlier training stage, the model mainly learns to recognize some 'easier-to-learn' discriminative patterns within each example, e.g., the lower-frequency components of images and the original information before data augmentation. Driven by this phenomenon, we propose a curriculum where the model always leverages all the training data at each epoch, while the curriculum starts with only exposing the 'easier-to-learn' patterns of each example, and introduces gradually more difficult patterns. To implement this idea, we 1) introduce a cropping operation in the Fourier spectrum of the inputs, which enables the model to learn from only the lower-frequency components efficiently, 2) demonstrate that exposing the features of original images amounts to adopting weaker data augmentation, and 3) integrate 1) and 2) and design a curriculum learning schedule with a greedy-search algorithm. The resulting approach, EfficientTrain, is simple, general, yet surprisingly effective. As an off-the-shelf method, it reduces the wall-time training cost of a wide variety of popular models (e.g., ResNet, ConvNeXt, DeiT, PVT, Swin, and CSWin) by >1.5x on ImageNet-1K/22K without sacrificing accuracy. It is also effective for self-supervised learning (e.g., MAE). Code is available at https://github.com/LeapLabTHU/EfficientTrain.
Satellite imagery is being leveraged for many societally critical tasks across climate, economics, and public health. Yet, because of heterogeneity in landscapes (e.g. how a road looks in different places), models can show disparate performance across geographic areas. Given the important potential of disparities in algorithmic systems used in societal contexts, here we consider the risk of urban-rural disparities in identification of land-cover features. This is via semantic segmentation (a common computer vision task in which image regions are labelled according to what is being shown) which uses pre-trained image representations generated via contrastive self-supervised learning. We propose fair dense representation with contrastive learning (FairDCL) as a method for de-biasing the multi-level latent space of convolution neural network models. The method improves feature identification by removing spurious model representations which are disparately distributed across urban and rural areas, and is achieved in an unsupervised way by contrastive pre-training. The obtained image representation mitigates downstream urban-rural prediction disparities and outperforms state-of-the-art baselines on real-world satellite images. Embedding space evaluation and ablation studies further demonstrate FairDCL's robustness. As generalizability and robustness in geographic imagery is a nascent topic, our work motivates researchers to consider metrics beyond average accuracy in such applications.
Children's speech recognition is a vital, yet largely overlooked domain when building inclusive speech technologies. The major challenge impeding progress in this domain is the lack of adequate child speech corpora; however, recent advances in self-supervised learning have created a new opportunity for overcoming this problem of data scarcity. In this paper, we leverage self-supervised adult speech representations and use three well-known child speech corpora to build models for children's speech recognition. We assess the performance of fine-tuning on both native and non-native children's speech, examine the effect of cross-domain child corpora, and investigate the minimum amount of child speech required to fine-tune a model which outperforms a state-of-the-art adult model. We also analyze speech recognition performance across children's ages. Our results demonstrate that fine-tuning with cross-domain child corpora leads to relative improvements of up to 46.08% and 45.53% for native and non-native child speech respectively, and absolute improvements of 14.70% and 31.10%. We also show that with as little as 5 hours of transcribed children's speech, it is possible to fine-tune a children's speech recognition system that outperforms a state-of-the-art adult model fine-tuned on 960 hours of adult speech.
Unsupervised Domain Adaptation (UDA) aims to enhance the generalization of the learned model to other domains. The domain-invariant knowledge is transferred from the model trained on labeled source domain, e.g., video game, to unlabeled target domains, e.g., real-world scenarios, saving annotation expenses. Existing UDA methods for semantic segmentation usually focus on minimizing the inter-domain discrepancy of various levels, e.g., pixels, features, and predictions, for extracting domain-invariant knowledge. However, the primary intra-domain knowledge, such as context correlation inside an image, remains underexplored. In an attempt to fill this gap, we propose a unified pixel- and patch-wise self-supervised learning framework, called PiPa, for domain adaptive semantic segmentation that facilitates intra-image pixel-wise correlations and patch-wise semantic consistency against different contexts. The proposed framework exploits the inherent structures of intra-domain images, which: (1) explicitly encourages learning the discriminative pixel-wise features with intra-class compactness and inter-class separability, and (2) motivates the robust feature learning of the identical patch against different contexts or fluctuations. Extensive experiments verify the effectiveness of the proposed method, which obtains competitive accuracy on the two widely-used UDA benchmarks, i.e., 75.6 mIoU on GTA to Cityscapes and 68.2 mIoU on Synthia to Cityscapes. Moreover, our method is compatible with other UDA approaches to further improve the performance without introducing extra parameters.
We present a self-supervised algorithm for several classification tasks within hematoxylin and eosin (H&E) stained images of breast cancer. Our method is robust to stain variations inherent to the histology images acquisition process, which has limited the applicability of automated analysis tools. We address this problem by imposing constraints a learnt latent space which leverages stain normalization techniques during training. At every iteration, we select an image as a normalization target and generate a version of every image in the batch normalized to that target. We minimize the distance between the embeddings that correspond to the same image under different staining variations while maximizing the distance between other samples. We show that our method not only improves robustness to stain variations across multi-center data, but also classification performance through extensive experiments on various normalization targets and methods. Our method achieves the state-of-the-art performance on several publicly available breast cancer datasets ranging from tumor classification (CAMELYON17) and subtyping (BRACS) to HER2 status classification and treatment response prediction.
Linguistic anomalies detectable in spontaneous speech have shown promise for various clinical applications including screening for dementia and other forms of cognitive impairment. The feasibility of deploying automated tools that can classify language samples obtained from speech in large-scale clinical settings depends on the ability to capture and automatically transcribe the speech for subsequent analysis. However, the impressive performance of self-supervised learning (SSL) automatic speech recognition (ASR) models with curated speech data is not apparent with challenging speech samples from clinical settings. One of the key questions for successfully applying ASR models for clinical applications is whether imperfect transcripts they generate provide sufficient information for downstream tasks to operate at an acceptable level of accuracy. In this study, we examine the relationship between the errors produced by several deep learning ASR systems and their impact on the downstream task of dementia classification. One of our key findings is that, paradoxically, ASR systems with relatively high error rates can produce transcripts that result in better downstream classification accuracy than classification based on verbatim transcripts.
In this paper, we provide a new perspective on self-supervised speech models from how the training targets are obtained. We generalize the targets extractor into Offline Targets Extractor (Off-TE) and Online Targets Extractor (On-TE). Based on this, we propose a new multi-tasking learning framework for self-supervised learning, MT4SSL, which stands for Boosting Self-Supervised Speech Representation Learning by Integrating Multiple Targets. MT4SSL uses the K-means algorithm as an Off-TE and a teacher network without gradients as an On-TE, respectively. Our model outperforms previous SSL methods by nontrivial margins on the LibriSpeech benchmark, and is comparable to or even better than the best-performing models with fewer data. Furthermore, we find that using both Off-TE and On-TE results in better convergence in the pre-training phase. With both effectiveness and efficiency, we think doing multi-task learning on self-supervised speech models from our perspective is a promising trend.
Self-supervised pre-training bears potential to generate expressive representations without human annotation. Most pre-training in Earth observation (EO) are based on ImageNet or medium-size, labeled remote sensing (RS) datasets. We share an unlabeled RS dataset SSL4EO-S12 (Self-Supervised Learning for Earth Observation - Sentinel-1/2) to assemble a large-scale, global, multimodal, and multi-seasonal corpus of satellite imagery from the ESA Sentinel-1 \& -2 satellite missions. For EO applications we demonstrate SSL4EO-S12 to succeed in self-supervised pre-training for a set of methods: MoCo-v2, DINO, MAE, and data2vec. Resulting models yield downstream performance close to, or surpassing accuracy measures of supervised learning. In addition, pre-training on SSL4EO-S12 excels compared to existing datasets. We make openly available the dataset, related source code, and pre-trained models at https://github.com/zhu-xlab/SSL4EO-S12.
In Self-Supervised Learning (SSL), various pretext tasks are designed for learning feature representations through contrastive loss. However, previous studies have shown that this loss is less tolerant to semantically similar samples due to the inherent defect of instance discrimination objectives, which may harm the quality of learned feature embeddings used in downstream tasks. To improve the discriminative ability of feature embeddings in SSL, we propose a new loss function called Angular Contrastive Loss (ACL), a linear combination of angular margin and contrastive loss. ACL improves contrastive learning by explicitly adding an angular margin between positive and negative augmented pairs in SSL. Experimental results show that using ACL for both supervised and unsupervised learning significantly improves performance. We validated our new loss function using the FSDnoisy18k dataset, where we achieved 73.6% and 77.1% accuracy in sound event classification using supervised and self-supervised learning, respectively.
In this paper, we introduce 3D-CSL, a compact pipeline for Near-Duplicate Video Retrieval (NDVR), and explore a novel self-supervised learning strategy for video similarity learning. Most previous methods only extract video spatial features from frames separately and then design kinds of complex mechanisms to learn the temporal correlations among frame features. However, parts of spatiotemporal dependencies have already been lost. To address this, our 3D-CSL extracts global spatiotemporal dependencies in videos end-to-end with a 3D transformer and find a good balance between efficiency and effectiveness by matching on clip-level. Furthermore, we propose a two-stage self-supervised similarity learning strategy to optimize the entire network. Firstly, we propose PredMAE to pretrain the 3D transformer with video prediction task; Secondly, ShotMix, a novel video-specific augmentation, and FCS loss, a novel triplet loss, are proposed further promote the similarity learning results. The experiments on FIVR-200K and CC_WEB_VIDEO demonstrate the superiority and reliability of our method, which achieves the state-of-the-art performance on clip-level NDVR.
This paper presents miCSE, a mutual information-based contrastive learning framework that significantly advances the state-of-the-art in few-shot sentence embedding. The proposed approach imposes alignment between the attention pattern of different views during contrastive learning. Learning sentence embeddings with miCSE entails enforcing the structural consistency across augmented views for every sentence, making contrastive self-supervised learning more sample efficient. As a result, the proposed approach shows strong performance in the few-shot learning domain. While it achieves superior results compared to state-of-the-art methods on multiple benchmarks in few-shot learning, it is comparable in the full-shot scenario. This study opens up avenues for efficient self-supervised learning methods that are more robust than current contrastive methods for sentence embedding.
Training robust speaker verification systems without speaker labels has long been a challenging task. Previous studies observed a large performance gap between self-supervised and fully supervised methods. In this paper, we apply a non-contrastive self-supervised learning framework called DIstillation with NO labels (DINO) and propose two regularization terms applied to embeddings in DINO. One regularization term guarantees the diversity of the embeddings, while the other regularization term decorrelates the variables of each embedding. The effectiveness of various data augmentation techniques are explored, on both time and frequency domain. A range of experiments conducted on the VoxCeleb datasets demonstrate the superiority of the regularized DINO framework in speaker verification. Our method achieves the state-of-the-art speaker verification performance under a single-stage self-supervised setting on VoxCeleb. Code has been made publicly available at https://github.com/alibaba-damo-academy/3D-Speaker.
Pre-training large transformer models with in-domain data improves domain adaptation and helps gain performance on the domain-specific downstream tasks. However, sharing models pre-trained on potentially sensitive data is prone to adversarial privacy attacks. In this paper, we asked to which extent we can guarantee privacy of pre-training data and, at the same time, achieve better downstream performance on legal tasks without the need of additional labeled data. We extensively experiment with scalable self-supervised learning of transformer models under the formal paradigm of differential privacy and show that under specific training configurations we can improve downstream performance without sacrifying privacy protection for the in-domain data. Our main contribution is utilizing differential privacy for large-scale pre-training of transformer language models in the legal NLP domain, which, to the best of our knowledge, has not been addressed before.
While deep reinforcement learning has proven to be successful in solving control tasks, the "black-box" nature of an agent has received increasing concerns. We propose a prototype-based post-hoc policy explainer, ProtoX, that explains a blackbox agent by prototyping the agent's behaviors into scenarios, each represented by a prototypical state. When learning prototypes, ProtoX considers both visual similarity and scenario similarity. The latter is unique to the reinforcement learning context, since it explains why the same action is taken in visually different states. To teach ProtoX about visual similarity, we pre-train an encoder using contrastive learning via self-supervised learning to recognize states as similar if they occur close together in time and receive the same action from the black-box agent. We then add an isometry layer to allow ProtoX to adapt scenario similarity to the downstream task. ProtoX is trained via imitation learning using behavior cloning, and thus requires no access to the environment or agent. In addition to explanation fidelity, we design different prototype shaping terms in the objective function to encourage better interpretability. We conduct various experiments to test ProtoX. Results show that ProtoX achieved high fidelity to the original black-box agent while providing meaningful and understandable explanations.
In recent years, research on visual document understanding (VDU) has grown significantly, with a particular emphasis on the development of self-supervised learning methods. However, one of the significant challenges faced in this field is the limited availability of publicly accessible visual corpora or extensive collections of images with detailed text annotations, particularly for non-Latin or resource-scarce languages. To address this challenge, we propose Web-based Visual Corpus Builder (Webvicob), a dataset generator engine capable of constructing large-scale, multilingual visual corpora from raw Wikipedia HTML dumps. Our experiments demonstrate that the data generated by Webvicob can be used to train robust VDU models that perform well on various downstream tasks, such as DocVQA and post-OCR parsing. Furthermore, when using a dataset of 1 million images generated by Webvicob, we observed an improvement of over 13% on the DocVQA Task 3 compared to a dataset of 11 million images from the IIT-CDIP. The implementation of our engine is publicly available on https://github.com/clovaai/webvicob
Modern speech enhancement (SE) networks typically implement noise suppression through time-frequency masking, latent representation masking, or discriminative signal prediction. In contrast, some recent works explore SE via generative speech synthesis, where the system's output is synthesized by a neural vocoder after an inherently lossy feature-denoising step. In this paper, we propose a denoising vocoder (DeVo) approach, where a vocoder accepts noisy representations and learns to directly synthesize clean speech. We leverage rich representations from self-supervised learning (SSL) speech models to discover relevant features. We conduct a candidate search across 15 potential SSL front-ends and subsequently train our vocoder adversarially with the best SSL configuration. Additionally, we demonstrate a causal version capable of running on streaming audio with 10ms latency and minimal performance degradation. Finally, we conduct both objective evaluations and subjective listening studies to show our system improves objective metrics and outperforms an existing state-of-the-art SE model subjectively.