Learning salient representations of multiview data is an essential step in many applications such as image classification, retrieval, and annotation. Standard predictive methods, such as support vector machines, often directly use all the features available without taking into consideration the presence of distinct views and the resultant view dependencies, coherence, and complementarity that offer key insights to the semantics of the data, and are therefore offering weak performance and are incapable of supporting view-level analysis. This paper presents a statistical method to learn a predictive subspace representation underlying multiple views, leveraging both multiview dependencies and availability of supervising side-information. Our approach is based on a multiview latent subspace Markov network (MN) which fulfills a weak conditional independence assumption that multiview observations and response variables are conditionally independent given a set of latent variables. To learn the latent subspace MN, we develop a large-margin approach which jointly maximizes data likelihood and minimizes a prediction loss on training data. Learning and inference are efficiently done with a contrastive divergence method. Finally, we extensively evaluate the large-margin latent MN on real image and hotel review datasets for classification, regression, image annotation, and retrieval. Our results demonstrate that the large-margin approach can achieve significant improvements in terms of prediction performance and discovering predictive latent subspace representations.

With the rapid development of deep sequencing techniques in the recent years, enhancers have been systematically identified in such projects as FANTOM and ENCODE, forming genome-wide landscapes in a series of human cell lines. Nevertheless, experimental approaches are still costly and time consuming for large scale identification of enhancers across a variety of tissues under different disease status, making computational identification of enhancers indispensable. To facilitate the identification of enhancers, we propose a computational framework, named DeepEnhancer, to distinguish enhancers from background genomic sequences. Our method purely relies on DNA sequences to predict enhancers in an end-to-end manner by using a deep convolutional neural network (CNN). We train our deep learning model on permissive enhancers and then adopt a transfer learning strategy to fine-tune the model on enhancers specific to a cell line. Results demonstrate the effectiveness and efficiency of our method in the classification of enhancers against random sequences, exhibiting advantages of deep learning over traditional sequence-based classifiers. We then construct a variety of neural networks with different architectures and show the usefulness of such techniques as max-pooling and batch normalization in our method. To gain the interpretability of our approach, we further visualize convolutional kernels as sequence logos and successfully identify similar motifs in the JASPAR database. DeepEnhancer enables the identification of novel enhancers using only DNA sequences via a highly accurate deep learning model. The proposed computational framework can also be applied to similar problems, thereby prompting the use of machine learning methods in life sciences.

Objective . Auditory attention decoding (AAD) determines which speaker the listener is focusing on by analyzing his/her EEG. Convolutional neural network (CNN) was adopted to extract spectro-spatial-feature (SSF) from short-time-interval of EEG to detect auditory spatial attention without stimuli. However, the following factors are not considered in SSF-CNN scheme. (a) Single-band frequency analysis cannot represent the EEG pattern precisely. (b) The power cannot represent the EEG feature related to the dynamic patterns of the attended auditory stimulus. (c) The temporal feature of EEG representing the relationship between EEG and attended stimulus is not extracted. To solve these problems, SSF-CNN scheme was modified. Approach . (a) Multiple-frequency bands, but not a single alpha frequency band, of EEG, were analyzed to represent the EEG pattern more precisely. (b) Differential entropy, but not power, was extracted from each frequency band to represent the disorder degree of EEG, which was related to the dynamic patterns of the attended auditory stimulus. (c) CNN and convolutional-long-short-term-memory (ConvLSTM) were combined to extract spectro-spatial-temporal features from the 3D descriptor sequence constructed based on the topographical activity maps of multiple-frequency bands. Main results . Experimental results on KUL, DTU, and PKU with 0.1 s, 1 s, 2 s, and 5 s decision windows demonstrated that: (a) The proposed model outperformed SSF-CNN and state-of-the-art AAD models. Specifically, when the auditory stimulus was unavailable, AAD accuracy could be enhanced by at least 3.25 % , 3.96 % and 5.08 % on KUL, DTU, and PKU, respectively, compared with the baselines. And, on KUL, the longer decision window corresponded to lower enhancement, while on both DTU and PKU, the longer decision window corresponded to higher enhancement, except for two cases when decision window length was 2 s on PKU or 5 s on DTU. (b) Each modification contributed to the performance enhancement. Significance . DE feature, multi-band frequency analysis, and ConvLSTM-based temporal analysis help to enhance AAD accuracy.

Objective . To take full advantage of both labeled data and unlabeled ones, the Graph Convolutional Network (GCN) was introduced in electroencephalography (EEG) based emotion recognition to achieve feature propagation. However, a single feature cannot represent the emotional state entirely and precisely due to the instability of the EEG signal and the complexity of the emotional state. In addition, the noise existing in the graph may affect the performance greatly. To solve these problems, it was necessary to introduce feature/similarity fusion and noise reduction strategies. Approach . A semi-supervised EEG emotion recognition model combining graph fusion, network enhancement, and feature fusion was proposed. Firstly, different features were extracted from EEG and then compacted by Principal Component Analysis (PCA), respectively. Secondly, a Sample-by-sample Similarity Matrix (SSM) was constructed based on each feature, and similarity network fusion (SNF) was adopted to fuse the graphs corresponding to different SSMs to take advantage of their complementarity. Then, Network Enhancement (NE) was performed on the fused graph to reduce the noise in it. Finally, GCN was performed on the concatenated features and the enhanced fused graph to achieve feature propagation. Main results . Experimental results demonstrated that: (a) When 5.30 % of SEED and 7.20 % of SEED-IV samples were chosen as the labeled samples, respectively, the minimum classification accuracy improvement achieved by the proposed scheme over state-of-the-art schemes were 1.52 % on SEED and 13.14 % on SEED-IV, respectively. (b) When 8.00 % of SEED and 9.60 % of SEED-IV samples were chosen as the labeled samples, respectively, the minimum training time reduction achieved by the proposed scheme over state-of-the-art schemes were 46.75 s and 22.55 s, respectively. (c) Graph fusion, network enhancement, and feature fusion all contributed to the performance enhancement. (d) The key hyperparameters that affect the performance were relatively few and easy to set to obtain outstanding performance. Significance . This paper demonstrated that the combination of graph fusion, network enhancement, and feature fusion help to enhance GCN-based EEG emotion recognition.