This paper introduces a human-in-the-loop (HITL) data annotation pipeline to generate high-quality, large-scale speech datasets. The pipeline combines human and machine advantages to more quickly, accurately, and cost-effectively annotate datasets with machine pre-labeling and fully manual auditing. Quality control mechanisms such as blind testing, behavior monitoring, and data validation have been adopted in the annotation pipeline to mitigate potential bias introduced by machine-generated labels. Our A/B testing and pilot results demonstrated the HITL pipeline can improve annotation speed and capacity by at least 80% and quality is comparable to or higher than manual double pass annotation. We are leveraging this scalable pipeline to create and continuously grow ultra-high volume off-the-shelf (UHV-OTS) speech corpora for multiple languages, with the capability to expand to 10,000+ hours per language annually. Customized datasets can be produced from the UHV-OTS corpora using dynamic packaging. UHV-OTS is a long-term Appen project to support commercial and academic research data needs in speech processing. Appen will donate a number of free speech datasets from the UHV-OTS each year to support academic and open source community research under the CC-BY-SA license. We are also releasing the code of the data pre-processing and pre-tagging pipeline under the Apache 2.0 license to allow reproduction of the results reported in the paper.
Real-world visual recognition problems often exhibit long-tailed distributions, where the amount of data for learning in different categories shows significant imbalance. Standard classification models learned on such data distribution often make biased predictions towards the head classes while generalizing poorly to the tail classes. In this paper, we present two effective modifications of CNNs to improve network learning from long-tailed distribution. First, we present a Class Activation Map Calibration (CAMC) module to improve the learning and prediction of network classifiers, by enforcing network prediction based on important image regions. The proposed CAMC module highlights the correlated image regions across data and reinforces the representations in these areas to obtain a better global representation for classification. Furthermore, we investigate the use of normalized classifiers for representation learning in long-tailed problems. Our empirical study demonstrates that by simply scaling the outputs of the classifier with an appropriate scalar, we can effectively improve the classification accuracy on tail classes without losing the accuracy of head classes. We conduct extensive experiments to validate the effectiveness of our design and we set new state-of-the-art performance on five benchmarks, including ImageNet-LT, Places-LT, iNaturalist 2018, CIFAR10-LT, and CIFAR100-LT.
Most of the few-shot learning methods learn to transfer knowledge from datasets with abundant labeled data (i.e., the base set). From the perspective of class space on base set, existing methods either focus on utilizing all classes under a global view by normal pretraining, or pay more attention to adopt an episodic manner to train meta-tasks within few classes in a local view. However, the interaction of the two views is rarely explored. As the two views capture complementary information, we naturally think of the compatibility of them for achieving further performance gains. Inspired by the mutual learning paradigm and binocular parallax, we propose a unified framework, namely Binocular Mutual Learning (BML), which achieves the compatibility of the global view and the local view through both intra-view and cross-view modeling. Concretely, the global view learns in the whole class space to capture rich inter-class relationships. Meanwhile, the local view learns in the local class space within each episode, focusing on matching positive pairs correctly. In addition, cross-view mutual interaction further promotes the collaborative learning and the implicit exploration of useful knowledge from each other. During meta-test, binocular embeddings are aggregated together to support decision-making, which greatly improve the accuracy of classification. Extensive experiments conducted on multiple benchmarks including cross-domain validation confirm the effectiveness of our method.
The instance discrimination paradigm has become dominant in unsupervised learning. It always adopts a teacher-student framework, in which the teacher provides embedded knowledge as a supervision signal for the student. The student learns meaningful representations by enforcing instance spatial consistency with the views from the teacher. However, the outputs of the teacher can vary dramatically on the same instance during different training stages, introducing unexpected noise and leading to catastrophic forgetting caused by inconsistent objectives. In this paper, we first integrate instance temporal consistency into current instance discrimination paradigms, and propose a novel and strong algorithm named Temporal Knowledge Consistency (TKC). Specifically, our TKC dynamically ensembles the knowledge of temporal teachers and adaptively selects useful information according to its importance to learning instance temporal consistency. Experimental result shows that TKC can learn better visual representations on both ResNet and AlexNet on linear evaluation protocol while transfer well to downstream tasks. All experiments suggest the good effectiveness and generalization of our method.
Few-shot object detection, which aims at detecting novel objects rapidly from extremely few annotated examples of previously unseen classes, has attracted significant research interest in the community. Most existing approaches employ the Faster R-CNN as basic detection framework, yet, due to the lack of tailored considerations for data-scarce scenario, their performance is often not satisfactory. In this paper, we look closely into the conventional Faster R-CNN and analyze its contradictions from two orthogonal perspectives, namely multi-stage (RPN vs. RCNN) and multi-task (classification vs. localization). To resolve these issues, we propose a simple yet effective architecture, named Decoupled Faster R-CNN (DeFRCN). To be concrete, we extend Faster R-CNN by introducing Gradient Decoupled Layer for multi-stage decoupling and Prototypical Calibration Block for multi-task decoupling. The former is a novel deep layer with redefining the feature-forward operation and gradient-backward operation for decoupling its subsequent layer and preceding layer, and the latter is an offline prototype-based classification model with taking the proposals from detector as input and boosting the original classification scores with additional pairwise scores for calibration. Extensive experiments on multiple benchmarks show our framework is remarkably superior to other existing approaches and establishes a new state-of-the-art in few-shot literature.
We address the challenging task of few-shot segmentation in this work. It is essential for few-shot semantic segmentation to fully utilize the support information. Previous methods typically adapt masked average pooling over the support feature to extract the support clues as a global vector, usually dominated by the salient part and loses some important clues. In this work, we argue that every support pixel's information is desired to be transferred to all query pixels and propose a Correspondence Matching Network (CMNet) with an Optimal Transport Matching module to mine out the correspondence between the query and support images. Besides, it is important to fully utilize both local and global information from the annotated support images. To this end, we propose a Message Flow module to propagate the message along the inner-flow within the same image and cross-flow between support and query images, which greatly help enhance the local feature representations. We further address the few-shot segmentation as a multi-task learning problem to alleviate the domain gap issue between different datasets. Experiments on PASCAL VOC 2012, MS COCO, and FSS-1000 datasets show that our network achieves new state-of-the-art few-shot segmentation performance.
Fundamental machine learning theory shows that different samples contribute unequally both in learning and testing processes. Contemporary studies on DNN imply that such sample di?erence is rooted on the distribution of intrinsic pattern information, namely sample regularity. Motivated by the recent discovery on network memorization and generalization, we proposed a pair of sample regularity measures for both processes with a formulation-consistent representation. Specifically, cumulative binary training/generalizing loss (CBTL/CBGL), the cumulative number of correct classi?cations of the training/testing sample within training stage, is proposed to quantize the stability in memorization-generalization process; while forgetting/mal-generalizing events, i.e., the mis-classification of previously learned or generalized sample, are utilized to represent the uncertainty of sample regularity with respect to optimization dynamics. Experiments validated the effectiveness and robustness of the proposed approaches for mini-batch SGD optimization. Further applications on training/testing sample selection show the proposed measures sharing the uni?ed computing procedure could benefit for both tasks.
In this work, we aim to address the challenging task of open set recognition (OSR). Many recent OSR methods rely on auto-encoders to extract class-specific features by a reconstruction strategy, requiring the network to restore the input image on pixel-level. This strategy is commonly over-demanding for OSR since class-specific features are generally contained in target objects, not in all pixels. To address this shortcoming, here we discard the pixel-level reconstruction strategy and pay more attention to improving the effectiveness of class-specific feature extraction. We propose a mutual information-based method with a streamlined architecture, Maximal Mutual Information Open Set Recognition (M2IOSR). The proposed M2IOSR only uses an encoder to extract class-specific features by maximizing the mutual information between the given input and its latent features across multiple scales. Meanwhile, to further reduce the open space risk, latent features are constrained to class conditional Gaussian distributions by a KL-divergence loss function. In this way, a strong function is learned to prevent the network from mapping different observations to similar latent features and help the network extract class-specific features with desired statistical characteristics. The proposed method significantly improves the performance of baselines and achieves new state-of-the-art results on several benchmarks consistently.
Unsupervised domain adaptive person re-identification (UDA re-ID) aims at transferring the labeled source domain's knowledge to improve the model's discriminability on the unlabeled target domain. From a novel perspective, we argue that the bridging between the source and target domains can be utilized to tackle the UDA re-ID task, and we focus on explicitly modeling appropriate intermediate domains to characterize this bridging. Specifically, we propose an Intermediate Domain Module (IDM) to generate intermediate domains' representations on-the-fly by mixing the source and target domains' hidden representations using two domain factors. Based on the "shortest geodesic path" definition, i.e., the intermediate domains along the shortest geodesic path between the two extreme domains can play a better bridging role, we propose two properties that these intermediate domains should satisfy. To ensure these two properties to better characterize appropriate intermediate domains, we enforce the bridge losses on intermediate domains' prediction space and feature space, and enforce a diversity loss on the two domain factors. The bridge losses aim at guiding the distribution of appropriate intermediate domains to keep the right distance to the source and target domains. The diversity loss serves as a regularization to prevent the generated intermediate domains from being over-fitting to either of the source and target domains. Our proposed method outperforms the state-of-the-arts by a large margin in all the common UDA re-ID tasks, and the mAP gain is up to 7.7% on the challenging MSMT17 benchmark. Code is available at https://github.com/SikaStar/IDM.
Hypergraph offers a framework to depict the multilateral relationships in real-world complex data. Predicting higher-order relationships, i.e hyperedge, becomes a fundamental problem for the full understanding of complicated interactions. The development of graph neural network (GNN) has greatly advanced the analysis of ordinary graphs with pair-wise relations. However, these methods could not be easily extended to the case of hypergraph. In this paper, we generalize the challenges of GNN in representing higher-order data in principle, which are edge- and node-level ambiguities. To overcome the challenges, we present SNALS that utilizes bipartite graph neural network with structural features to collectively tackle the two ambiguity issues. SNALS captures the joint interactions of a hyperedge by its local environment, which is retrieved by collecting the spectrum information of their connections. As a result, SNALS achieves nearly 30% performance increase compared with most recent GNN-based models. In addition, we applied SNALS to predict genetic higher-order interactions on 3D genome organization data. SNALS showed consistently high prediction accuracy across different chromosomes, and generated novel findings on 4-way gene interaction, which is further validated by existing literature.