Temporal relation classification is becoming an active research field. Lots of methods have been proposed, while most of them focus on extracting features from external resources. Less attention has been paid to a significant advance in a closely related task: relation extraction. In this work, we borrow a state-of-the-art method in relation extraction by adopting bidirectional long short-term memory (Bi-LSTM) along dependency paths (DP). We make a "common root" assumption to extend DP representations of cross-sentence links. In the final comparison to two stateof-the-art systems on TimeBank-Dense, our model achieves comparable performance, without using external knowledge and manually annotated attributes of entities (class, tense, polarity, etc.).

The Gaussian process (GP) approaches to classification synthesize Bayesian methods and kernel techniques, which are developed for the purpose of small sample analysis. Here we propose a GP model and investigate it for the facial expression recognition in the Japanese female facial expression dataset. By the strategy of leave-one-out cross validation, the accuracy of the GP classifiers reaches 93.43% without any feature selection/extraction. Even when tested on all expressions of any particular expressor, the GP classifier trained by the other samples outperforms some frequently used classifiers significantly. In order to survey the robustness of this novel method, the random trial of 10-fold cross validations is repeated many times to provide an overview of recognition rates. The experimental results demonstrate a promising performance of this application.

Object detection is an important component of computer vision. Most of the recent successful object detection methods are based on convolutional neural networks (CNNs). To improve the performance of these networks, researchers have designed many different architectures. They found that the CNN performance benefits from carefully increasing the depth and width of their structures with respect to the spatial dimension. Some researchers have exploited the cardinality dimension. Others have found that skip and dense connections were also of benefit to performance. Recently, attention mechanisms on the channel dimension have gained popularity with researchers. Global average pooling is used in SENet to generate the input feature vector of the channel-wise attention unit. In this work, we argue that channel-wise attention can benefit from both global average pooling and global max pooling. We designed three novel attention units, namely, an adaptive channel-wise attention unit, an adaptive spatial-wise attention unit and an adaptive domain attention unit, to improve the performance of a CNN. Instead of concatenating the output of the two attention vectors generated by the two channel-wise attention sub-units, we weight the two attention vectors based on the output data of the two channel-wise attention sub-units. We integrated the proposed mechanism with the YOLOv3 and MobileNetv2 framework and tested the proposed network on the KITTI and Pascal VOC datasets. The experimental results show that YOLOv3 with the proposed attention mechanism outperforms the original YOLOv3 by mAP values of 2.9 and 1.2% on the KITTI and Pascal VOC datasets, respectively. MobileNetv2 with the proposed attention mechanism outperforms the original MobileNetv2 by a mAP value of 1.7% on the Pascal VOC dataset.

No abstract

We propose an AdversariaL training algorithm for commonsense InferenCE (ALICE). We apply small perturbations to word embeddings and minimize the resultant adversarial risk to regularize the model. We exploit a novel combination of two different approaches to estimate these perturbations: 1) using the true label and 2) using the model prediction. Without relying on any human-crafted features, knowledge bases or additional datasets other than the target datasets, our model boosts the finetuning performance of RoBERTa, achieving competitive results on multiple reading comprehension datasets that require commonsense inference.