Motor imagery brain‐computer interfaces (BCIs) have demonstrated great potential and attract world‐spread attentions. Due to the nonstationary character of the motor imagery signals, costly and boring calibration sessions must be proceeded before use. This prevents them from going into our realistic life. In this paper, the source subject’s data are explored to perform calibration for target subjects. Model trained on source subjects is transferred to work for target subjects, in which the critical problem to handle is the distribution shift. It is found that the performance of classification would be bad when only the marginal distributions of source and target are made closer, since the discriminative directions of the source and target domains may still be much different. In order to solve the problem, our idea comes that joint distribution adaptation is indispensable. It makes the classifier trained in the source domain perform well in the target domain. Specifically, a measure for joint distribution discrepancy (JDD) between the source and target is proposed. Experiments demonstrate that it can align source and target data according to the class they belong to. It has a direct relationship with classification accuracy and works well for transferring. Secondly, a deep neural network with joint distribution matching for zero‐training motor imagery BCI is proposed. It explores both marginal and joint distribution adaptation to alleviate distribution discrepancy across subjects and obtain effective and generalized features in an aligned common space. Visualizations of intermediate layers illustrate how and why the network works well. Experiments on the two datasets prove the effectiveness and strength compared to outstanding counterparts.

Applying convolution methods to domains that lack regular underlying structures is a challenging task for 3D vision. Existing methods require the manual design of feature representations suitable for the task or full-voxel-level analysis, which is memory intensive. In this paper, we propose a novel feature extraction method to facilitate 3D nonrigid shape analysis. Our approach, called 3D-MConv, extends convolution operations from regular grids to irregular mesh sets by parametrizing a series of convolutional templates and adopts a novel local perspective to ensure that the algorithm is more invariant against global isometric deformation and articulation. We carefully design the convolutional template as a polynomial function that flexibly represents the local shape. An unsupervised learning method is adopted to learn the convolutional template function. By using the convolution operation and the movement of the template on the model surface, we can obtain the distribution of the typical template shapes. We combine this distribution feature with the spatial co-occurrence information of typical template shapes modelled by Markov chains to form a high-level descriptor of a 3D model. The support vector machine method is used to classify the nonrigid 3D objects. Experiments on SHREC10 and SHREC15 demonstrate that 3D-MConv achieves state-of-the-art accuracy on standard benchmarks.

Multimanifold clustering separates data points approximately lying on a union of submanifolds into several clusters. In this paper, we propose a new nonparametric Bayesian model to handle the manifold data structure. In our framework, we first model the manifold mapping function between Euclidean space and topological space by applying a deep neural network, and then construct the corresponding generation process of multiple manifold data. To solve the posterior approximation problem, in the optimization procedure, we apply a variational auto-encoder-based optimization algorithm. Especially, as the manifold algorithm has poor performance on the real dataset where nonmanifold and manifold clusters are appearing simultaneously, we expand our proposed manifold algorithm by integrating it with the original Dirichlet process mixture model. Experimental results have been carried out to demonstrate the state-of-the-art clustering performance.

Common Spatial Pattern (CSP) has been proved to be one of the most efficient featureextracting methods for Brain Computer Interfaces (BCI), especially for motor imagery BCI. However, CSP is a supervised method and performs poorly when there are not enough labeled data. This paper aims to construct a minimum-training BCI, which means there are only a few labeled data, even none labeled data for target subjects. Under this condition, conventional CSP cannot work well. Therefore, source data (related labeled data from other subjects) are exploited and common filters across subjects are obtained using a clustering method. After that features are extracted and a semi-supervised support vector machine which transfers knowledge across subjects is proposed. Experiments illustrate the effectiveness of our algorithm. When there are none labeled data for target subjects, our algorithm outperforms two state of art algorithms in semi-supervised learning field, and as the amount of unlabeled data for target subjects become larger, the performance of our algorithm grows better and better, which is suitable for online use. When the amount of labeled data for target subjects (denoted as M in this paper) is small, our algorithm also shows its strength compared with corresponding outstanding algorithms. For dataset IVa of BCI competition III, our algorithm performs the best for all the subjects excluding ‘aw’ when M=20. Compared with counterparts, our method outperforms 6.6% for ‘aa’, 19.3% for ‘al’, 11.8% for ‘av’, 9.7% for ‘ay’ on average, respectively. When M=40, our method performs the best for ‘al’, ‘av’. It averagely outperforms 8.0% for ‘al’, 15.3% for ‘av’, respectively. For GigaDataset, averagely our method outperforms 4.4% for ‘sbj2’, 5.0% for ‘sbj4’, 7.1% for ‘sbj5’ when M=20, respectively. When M=40, our algorithm performs the best only for ‘sbj5’. It outperforms 8.6% averagely compared with counterparts. Although performances of our method are not the best for all the conditions, its performances are very robust and competitive.

Deep spectral clustering is a popular and efficient algorithm in unsupervised learning. However, deep spectral clustering methods are organized into three separate steps: affinity matrix learning, spectral embedding learning, and K-means clustering on spectral embedding. In this case, although each step can achieve its own performance, it is still difficult to obtain robust clustering results. In this letter, we propose a one-step adaptive spectral clustering network to overcome the aforementioned shortcomings. The network embeds the three parts of affinity matrix learning, spectral embedding learning, and indicator learning into a unified framework. The affinity matrix is adaptively adjusted by spectral embedding in a deep subspace. We introduce spectral rotation to discretize spectral embedding, which makes the spectral embedding and indicator be learned simultaneously to improve clustering quality. Each part of the model can be iteratively updated based on other parts to optimize the clustering results. Experimental results on four real datasets show the effectiveness of our method on the ACC and NMI clustering evaluation metrics. In particular, our method achieves an NMI of 0.932 and an ACC of 0.973 on the MNIST dataset, a decent performance boost compared to the best baseline.

In this paper we propose an approach for Chinese question analysis and answer extraction. A general question analysis process contains keyword extraction and question classification. Question classification plays a crucial role in automatic question answering. To implement the question classification, we have carried out experiments with Support Vector Machines (SVM) using four kinds of features: words, part of speech (POS), named entity, semantic. And the answer extraction is converted into a clustering problem. The experimental results show the excellent performance of the proposed approach.