• Pytorch code for the ZEN method
    This is an unsupervised pretraining method for Chinese that can take advantage of n-gram (multi-character) information, achieving better results than BERT.

  • Pytorch code for the GULF method
    This is a Neural Network training method with the aim of improving generalization error.

  • Pytorch code for the CFG-GAN method
    This is a method of more stable training of GAN.

  • C++ code implementaiton of various text catgorization methods using neural networks
         CNN (NAACL 15)Semisupervised Learning (NIPS 15) LSTM (ICML 16),  and DPCNN (ACL 17).

  • C++ code for nonlinear prediction using the RGF method
    This is a regularized version of boosted decision trees (forest) with improved performance.

  • C++ code for solving large scale L1-L2 regularization problems using the Prox-SDCA method
    This is an accelerated stochastic method to train finite sum linear models with fast convergence rate.