Search

Commonly used clustering algorithms typically require user parameters such as the number of clusters to be divided. Density-based algorithms do not have such requirements but are not suitable for high dimensional data. Recent studies have merged the cluster assignment task with deep similarity learning. In this paper, we propose a novel framework to perform dynamic image clustering without prior knowledge of the cluster count. A deep learning model first learns data similarity from scratch, followed by the use of a coordinate learning model to project high dimensional data onto a two-dimensional space. A new clustering algorithm, raster clustering, is proposed to evaluate and classify the projected data. This mechanism can be applied in high dimensional data clustering like image data, and it allows the prediction of unseen data in a consistent way without the need for consolidating with training data.

In brachytherapy, the segmentation accuracy of the target tumor and surrounding organs is very important. In recent years, deep learning models have improved the performance of organ segmentation and have been widely used. However, it is still a huge challenge for some organs with variable shapes. The basic idea of the work presented in this paper is to accurately divide the rectal computed tomography image dataset so that patients can obtain more accurate brachytherapy. In this work, we used 3D U-Net and Long ShortTerm Memory (LSTM) to improve the accuracy of rectal segmentation. This model was trained and tested on the rectal computed tomography image dataset, which contains 51 patients undergoing radiation therapy. The dice coefficient is used as the evaluation index in all results of organ segmentation. After experiments are done, it can be seen that the proposed method has good performance in rectal segmentation.

With the rising demand of learning neural network technology and analysis in recent years, how to diffuse the knowledge from sophisticated researchers to the general audience has become an essential issue. We have conducted a short survey on neural network visualization. Regarding the problem that existing tools for the neural network development provide limited visualization for model editing, we have implemented an easy-to-use neural network development platform prototype with full visualization on both model designing and results analysis. Users can now design, edit and analyze in a user-friendly browser-based environment, while training the model on one-click running through the web-generated program.

Rainfall nowcasting provides the estimations of rainfall condition, such as accumulated precipitation, probability of precipitation forecast, and rainfall intensity prediction. Although numerical weather prediction (NWP) can simulate the atmospheric conditions, limited by the computation performance and the initial field data, the NWP does not perform well in short-term forecasting. Since atmosphere environment is a complex non-linear system, we used the deep learning approach to learn and perform the rainfall nowcasting. In this paper, we used the classification model based on a residual network and added the "side path" to input the additional data which could assist our model in acquiring prior knowledge. For the experiment, we input the topographic data to help the model include topographical awareness. In our experiment, the model trained by the additional topographic data achieved the higher accuracy than the model lacking the topographical recognition.

In this paper we discuss how to determine the appropriate order when recommending new friends or new interests based on the confidence scores generated by using SVM on a social network. The main idea is analyzing the ego's friendship circles to find interests or friends which are not currently included in their profile, and then the confidence scores obtained by SVM are used to decide which interests or friends should be recommended in every turn, based on the focal individual's network structure, also called an ego network. The novel factor in this approach is that it does not make recommendations using a fixed order. This is because a fixed order would make some improper recommendations, then mislead next recommendation, and obtain increasingly less accuracy results. In our experiments, it is shown that the accuracy achieved when using our confidence score order is better than that with a fixed order, even when there is very little information contained in a person's profile.

The relationship between a personi¦s personality and performance has long been studied by psychologists. Research suggests that a personi¦s performance and behavior are related to personality characteristics and background data to a certain degree. In this paper, the Big Five personality model is adopted for measuring profiles of students, whose undergraduate performance and behavior are then analyzed. A machine learning approach, support vector regression (SVR), is employed to find correlations from the given sample data. The performance and behavior of a person are predicted from the obtained regression values. Personality, biological, performance, and behavior data of 120 undergraduates in Taiwan were collected through questionnaires. Ninety valid data samples are used for training in SVR and the others are used for evaluating the regression predictions. Most of the predicted performance yielded near 80% accuracy. It is shown that there are correlations between a personi¦s performance and personality characteristics. SVR is shown to be a suitable method for exploring personality correlations.