Jianjun Xu

Haoyu Zhang, Jianjun Xu, Ji Wang

In this paper, we propose a novel pretraining-based encoder-decoder framework, which can generate the output sequence based on the input sequence in a two-stage manner. For the encoder of our model, we encode the input sequence into context representations using BERT. For the decoder, there are two stages in our model, in the first stage, we use a Transformer-based decoder to generate a draft output sequence. In the second stage, we mask each word of the draft sequence and feed it to BERT, then by combining the input sequence and the draft representation generated by BERT, we use a Transformer-based decoder to predict the refined word for each masked position. To the best of our knowledge, our approach is the first method which applies the BERT into text generation tasks. As the first step in this direction, we evaluate our proposed method on the text summarization task. Experimental results show that our model achieves new state-of-the-art on both CNN/Daily Mail and New York Times datasets.

Zhen Li, Jianjun Xu, Tingshao Zhu

With the development of the modern society, mind control applied to both the recovery of disabled individuals and auxiliary control of normal people has obtained great attention in numerous researches. In our research, we attempt to recognize the brain waves of left and right hand movement imagery with portable electroencephalographs. Considering the inconvenience of wearing traditional multiple-electrode electroencephalographs, we choose Muse to collect data which is a portable headband launched lately with a number of useful functions and channels and it is much easier for the public to use. Additionally, previous researches generally focused on discrimination of EEG of left and right hand movement imagery by using data from C3 and C4 electrodes which locate on the top of the head. However, we choose the gamma wave channels of F7 and F8 and obtain data when subjects imagine their left or right hand to move with their eyeballs rotated in the corresponding direction. With the help of the Common Space Pattern algorithm to extract features of brain waves between left and right hand movement imagery, we make use of the Support Vector Machine to classify different brain waves. Traditionally, the accuracy rate of classification was approximately 90% using the EEG data from C3 and C4 electrode poles; however, the accuracy rate reaches 95.1% by using the gamma wave data from F7 and F8 in our experiment. Finally, we design a plane program in Python where a plane can be controlled to go left or right when users imagine their left or right hand to move. 8 subjects are tested and all of them can control the plane flexibly which reveals that our model can be applied to control hardware which is useful for disabled individuals and normal people.

Zhen Li, Jianjun Xu, Tingshao Zhu

In our research, we attempt to help people recognize their brain state of concentration or relaxation more conveniently and in real time. Considering the inconvenience of wearing traditional multiple electrode electroencephalographs, we choose Muse to collect data which is a portable headband launched lately with a number of useful functions and channels and it is much easier for the public to use. Besides, traditional online analysis did not focus on the synchronism between users and computers and the time delay problem did exist. To solve the problem, by building the Analytic Hierarchy Model, we choose the two gamma wave channels of F7 and F8 as the data source instead of using both beta and alpha channels traditionally.Using the Common Space Pattern algorithm and the Support Vector Machine model, the channels we choose have a higher recognition accuracy rate and smaller amount of data to be dealt with by the computer than the traditional ones. Furthermore, we make use of the Feedforward Neural Network Model to predict subjects'brain states in half a second. Finally, we design a plane program in Python where a plane can be controlled to go up or down when users concentrate or relax. The SVM model and the Feedforward Neural Network model have both been tested by 12 subjects and they give an evaluation ranging from 1 to 10 points. The former gets 7.58 points while the latter gets 8.83, which proves that the time delay problem is improved once more.