: In recent years, deep neural network has continuously renewed its best performance in tasks such as computer vision and natural language processing, and has become the most concerned research direction. Although the performance of deep network model is remarkable, it is still difficult to deploy to the embedded or mobile devices with a limited hardware due to the large number of parameters, high storage and computing costs. It has been found by relevant studies that the depth model based on convolutional neural network has parameter redundancy, and there are parameters that are useless to the final result in the model, which provides theoretical support for the pruning of depth network model. Therefore, how to reduce the model size under the condition of ensuring the model accuracy has become a hot issue. This paper classifies and summarizes the achievements of domestic and foreign scholars in model pruning in recent years, selects several new pruning algorithm methods in different directions, analyzes their functionality through experiments and discusses the current problems of different models and the development direction of pruning model optimization in the future.
In this research, we design an application to release user’s daily stress. It is said that contemporary people are in a pretty stressful situation. For example, in recent years, we have suffered from difficulty to meet friends casually and enjoy talking with them to get rid of our daily stress due to the COVID-19 pandemic. In other words, the variation of ways of overcoming stress have been limited. In addition, people have deteriorated relationships with their family members because of the longer stay-at-home time than before. Health problems such as the onset of depression caused by the stress of the pandemic are also becoming more serious. Moreover, people are worrying of increasing the number of suicides. This situation is one of the problems that people consider it important in the society recently. Therefore, there is an urgent need to find solutions to such a situation. In this research, we develop the functions to cope with daily stress easily for the application of the smartphones that anyone has. We include functions in this application we suggest in this research. They are “vent mode”, which is a dynamic de-stressing method by words, “empathy mode”, which is a static de-stressing method, and “physical touch mode”, which is a dynamic de-stressing method that is accompanied with physical exercise. In addition, we design the application for de-stress in order to break through the stressful situation these days.
: This paper introduces a novel technique to generate brand personalities for organizations from social media text data using a deep bi - directional Long Short - Term Memory (BiLSTM) neural network model in combination with an accepted brand personality scale model. Brand Personality has evolved into an indispensable element in modern business organizations. Currently brand personality of an organization is generated using traditional techniques such as stakeholder interviews, questionnaires, which are time and resource intensive and limited in efficacy. However, the rise of the internet and social media have provided a platform for stakeholders to frequently and freely express their opinions and experiences related to organizations. Such social media data while now successfully being used for customer analytics have not yet been fully investigated and used to understand brand personalities. Our research investigated how this data can be effectively leveraged by organizations to generate and monitor their brand in near real time. Our technique has been successfully demonstrated on Higher Education Institutes, as Higher Education is an industry that is increasingly being exposed to business competition over the last few decades.
The detection and recognition of information in natural scenes has always been a difficult problem in computer vision. Digital instrument character recognition is one of the more representative and valuable things. In recent years, there is a lot of research work on this problem, but the solutions rely on string location, character segmentation and other preprocessing processes, the results of these preprocessing processes directly affect the final character recognition results. In contrast, the character recognition method of digital instrument based on convolution neural network (CNN) omits the complex preprocessing process through graph to graph prediction, and the character recognition result is obtaind directly. And has a strong generalization ability, can identify multiple types of instruments. At the same time, through the weighted fusion of muti-scale and multi-level features in the CNN, a better ability of feature extraction and information integration is obtained. The experimental results show that the method can directly and accurately recognize the characters in the Vernier caliper.
A covert channel is an information channel which is used by computer process to exfiltrate data through bypassing security policies. The domain name system (DNS) protocol is one of the important ways to implement a covert channel. DNS covert channels are easily used by attackers for malicious purposes. Therefore, an effective detection of the DNS covert channels is significant for computer system and network security. Aiming at the difficulty of the DNS covert channel identification, we propose a DNS covert channel detection method based on stacking model. The stacking model is evaluated in a campus network and the experimental results show that the detection based on the stacking model can detect the DNS covert channels effectively. Besides, it can also identify unknown covert channel traffic. The area under the curve (AUC) of the proposed method, reaching 0.9901, outperforms the existed methods.
