Yue Zhou

Applied the Involution operator to the 1-dimensional data which originally applied to the 2-dimensional image, making the feature extraction operation lighter and more efficient than when using convolution. Used the Inception network to extract features from the data within 60 minutes, and constructed an end-to-end architecture, which greatly increases the amount of information input to the Transformer-based prediction model. Applied the inception network with involution operator to the Transformer-based prediction model. Using the same data set, the improved model improved the indicators of 64.4% of the model currently in use online. Applied the BERT structure to the field of time series analysis in order to achieve higher prediction accuracy and faster model training speed, and adjusted the masked LM in the BERT pre-training step to make it more suitable for time series analysis.

Added a new field to the pipeline of Cortana’s usage data and improved the Cortana Metrics dashboards. Applied a rule-based model to classify test users and achieved 81.8% accuracy and a 98% recall rate, respectively. Employed the theory of SVM to optimize the rule-based model and applied this model to the assistant data pipeline.

Conducted experiments on datasets of various sizes to verify the higher accuracy of the proposed method and compared it with Common Feature Learning (CFL). Trained teacher-related models with four different architecture combinations from pre-training models and improved the accuracy to 83.31%. Co-authored the paper Cross Attention on Hierarchical Feature Learning for Unsupervised Knowledge Amalgamation, which was submitted to the AAAI-21 Conference and reached Phase II of review process.

Conducted research on the CycleGAN theory and applied it to specific image reconstruction design. Implemented CycleGAN on different scenarios and compared their results, verifying that CycleGAN can achieve reliable performances in different application scenarios.