Search

Dysarthria is a manifestation of disruption in the neuromuscular physiology resulting in uneven, slow, slurred, harsh, or quiet speech. Despite the remarkable progress of automatic speech recognition (ASR), it poses great challenges in developing stable ASR for dysarthric individuals due to the high intra- and inter-speaker variations and data deficiency. In this paper, we propose a contrastive learning framework for robust dysarthric speech recognition (DSR) by capturing the dysarthric speech variability. Several speech data augmentation strategies are explored to form two branches of the framework, meanwhile alleviating the scarcity of dysarthria data. We also develop an efficient projection head acting on a sequence of learned hidden representations for defining contrastive loss. Experiment results on DSR demonstrate that the model is better than or comparable to the supervised baseline.

In multiview settings, observations from different views are assumed to share the same subspace. The abundance of views can be utilized to better explore the subspace. In this chapter, we consider two different kinds of multiview subspace learning problems. The first one contains the general unsupervised multiview subspace learning problems. We focus on canonical correlation analysis as well as some of its extensions. The second one contains the supervised multiview subspace learning problems, i.e., there exists available label information. In this case, representations more suitable for the on-hand task can be obtained by utilizing the label information. We also briefly introduce some other methods at the end of this chapter.

Transfer learning is proposed to transfer the learned knowledge from source domains to target domains where the target ones own fewer training data. Multitask learning learns multiple tasks simultaneously and makes use of the relationship among these tasks. Both of these learning methods can combine with the multiview learning, which exploits the information from the consistency of diverse views. In this chapter, we introduce four multiview transfer learning methods and three multiview multitask learning methods. We review research on multiview transfer learning under the large margin framework, discuss multiview discriminant transfer learning in detail, and introduce how to adapt Adaboost into multiview transfer learning. Three multiview multitask learning methods concentrate on the shared structures between tasks and views. The most natural way is to represent the relationships based on the bipartite graph and use an iterative algorithm to optimize its objective function. Another method constructs additional regularization function to ensure the view consistency. In general, convex shared structure learning algorithm provides structure parameters to share information. Besides, we introduce other methods; as supplements, where multi-transfer, multitask multiview discriminant analysis, and clustering are briefly mentioned.

Active learning is proposed based on the fact that manually labeled examples are expensive, thus it picks the most informative points to label so as to improve the learning efficiency. Combined with multiview learning algorithm, it constructs multiple learners to select contention points among different views. In this chapter, we introduce five multiview active learning algorithms as examples. At first, we introduce co-testing, the first algorithm applying active learning to multiview learning, and discuss how to process the contradiction between multiple learners. Bayesian co-training is proposed under the mutual information framework, which considers the unobserved labels as latent variables and marginalizes them out. We focus on multiview multi-learner learning active learning, which introduces the ambiguity of an example to measure its confidence. In the situation that active learning with extremely sparse labeled examples, there is a detailed derivation of CCA in two view. At last, we retell a practical active learning algorithm combined with semi-supervised learning. Besides, there are other methods briefly mentioned at the end of this chapter.

Semi-supervised learning is concerned with such learning scenarios where only a small portion of training data are labeled. In multiview settings, unlabeled data can be used to regularize the prediction functions, and thus to reduce the search space. In this chapter, we introduce two categories of multiview semi-supervised learning methods. The first one contains the co-training style methods, where the prediction functions from different views are trained through their own objective, and each prediction function is improved by the others. The second one contains the co-regularization style methods, where a single objective function exists for the prediction functions from different views to be trained simultaneously.

Multiview supervised learning algorithm can exploit the multiview nature of the data by the consensus of the views, that is, to seek predictors from different views that agree on the same example. In this chapter, we introduce three categories of multiview supervised learning methods. The first one contains the multiview large margin-based classifiers, which regularize the classifiers from different views with their agreements on classification margins to enforce view consensus. The second one contains multiple kernel learning, where the feature mappings underlying multiple kernels will map the views to new feature spaces where the classifiers are more likely to agree on the views. The third one contains some Gaussian process related models, in which case the predict functions themselves are taken as random variables. We also briefly introduce some other methods at the end of this chapter.

The multiview deep learning described in this chapter deals with multiview data or simulates constructing its intrinsic structure by using deep learning methods. We highlight three major categories of multiview deep learning methods through three different thoughts. The first category of approaches focuses on obtaining a shared joint representation from different views by building a hierarchical structure. The second category of approaches focuses on constructing structured spaces with different representations of multiple views which gives some constraints between representations on a different view. The third major category approaches focuses on explicitly constructing connections or relationships between different views or representations, which allows different views to be translated or mapped to each other.

For geostationary meteorological satellite (GSMS) remote sensing image registration, high computational cost and matching error are the two main challenging problems. To address these issues, this paper proposes a novel algorithm named slope-restricted multi-scale feature matching. In multi-scale feature matching, images are subsampled to different scales. From a small scale to a large scale, the offsets between the matched pairs are used to narrow the searching area of feature matching for the next larger scale. Thus, the feature matching is accomplished from coarse to fine, which will make the matching process more accurate and reduce errors. To enhance the matching performance, the outliers in the matched pairs are rectified by using slope-restricted rectification, which is based on local geometric similarity. Compared with other algorithms, the experimental results show that our proposed method is more accurate and efficient.