A Novel Teacher-Student Network for Sentiment Classification
PROCEEDINGS OF THE 2016 2ND INTERNATIONAL CONFERENCE ON ARTIFICIAL INTELLIGENCE AND INDUSTRIAL ENGINEERING (AIIE 2016)
Authors: Chen, Huajie; Wang, Eric Ke; Li, Feng; Yu, Wenli
Abstract
Compared with traditional text classification, many sentiments online such as product reviews are not standard, which are concise with clear standpoints. Researchers on sentiment classification face tremendous challenges. Although various sentiment analysis systems are available, they have many operation restrictions and are still far from perfect. In this paper, we propose a novel approach, Teacher-Student Network (TSN), for automatically classifying the sentiment of reviews. Teacher-Student Network Model is composed of one teacher network and one student network. Teacher network is a Naive Bayes model. Student network is deep neural networks model. Our approach can transfer knowledge between different models and requires less training data. Experimental results on different domain datasets show that when we employ full training data, our model can achieve similar performance to RNN(Recurrent Neural Network) model and when we reduce training data, our model achieve better performance than RNN.
A Hardware/Software Co-Design Approach for Ethernet Controllers to Support Time-triggered Traffic in the Upcoming IEEE TSN Standards
2014 IEEE FOURTH INTERNATIONAL CONFERENCE ON CONSUMER ELECTRONICS BERLIN (ICCE-BERLIN)
Authors: Gross, Friedrich; Steinbach, Till; Korf, Franz; Schmidt, Thomas C.; Schwarz, Bernd
Abstract
Due to the increasing bandwidth and timing requirements, next generation communication backbones in cars will most likely base on real-time Ethernet variants that satisfy the demands of the new automotive applications. The upcoming IEEE 802.1Qbv standard shows communication approaches based on coordinated time devision multiple access (TDMA) to be good candidates for providing communication with determinism and highly precise timing. Implementing time-triggered architectures in software requires significant development effort and computational power. This paper shows a scalable HW/SW co-design approach for new real-time Ethernet controllers based on the partitioning into communication and application components. The tasks required for communication are divided: Time-critical and computationally intensive parts are realised in dedicated hardware modules allowing the attached CPU to fulfil the timing requirements of the automotive application without interference. The evaluation using a Field Programmable Gate Array (FPGA) based prototype implementation shows that the precision for the time-triggered transmission and the performance of the proposed implementation of the required synchronisation protocols satisfies the requirements of applications in the automotive domain.