Improving Question Answering with Knowledge Graphs
NLP Course Project at UCL, Feb 2019 - Mar. 2019.
Inspired by (Zhong et al., 2018), we propose to extract concept graphs from sentences and use the graph structures to answer commonsense-based questions. Our method obtained an accuracy of 62.56% (an absolute improvement of 1.05% over the vanilla BERT-based model). We also introduce some handcrafted rules to reformulate questions, which results in a higher accuracy of 63.93%.
Human Activity Recognition with SVM
Numerical Optimization Course Project at UCL, May 2019 - May 2019.
In this project, sequential minimal optimization and augmented Lagrangian method are implemented to solve the dual problem of SVM. Human Activity Recognition Dataset is used to test multi-class kernel SVM performance and different kernels, optimization methods, multi-class methods, and regularization are studied.
Playing Atari with Deep Reinforcement Learning
Research Project at Ecole Polytechnique, Oct. 2017 - Mar. 2018.
This project is mainly based on (Oh et al., 2015). I reimplemented the algorithm using Keras and Tensorflow and tested multiple different research ideas on auto-encoder. Some studies are also carried out on VAE and A3C.
Link Prediction in Citation Networks
NLP Course Project at Ecole Polytechnique, Feb. 2018 - Mar. 2018.
In this project, we extract both structure-based (neighbourhood, node2vec) and node-based (word2vec, doc2vec) features to perform link prediction in citation networks. Different neural network as well as tree-based algorithms (Xgboost, LightGBM) are compared.
Hacking the Paris metro
Big Data Course Project at Ecole Polytechnique, May 2017 - Jun. 2017.
Basic machine learning algorithms (SVM, Decision Tree, etc.) are applied to predict the train delay using real data acquired from RATP.
Lotus Effect and Spinodal Decomposition Phenomenon
Numerical Simulation of PDE Course Project at Ecole Polytechnique, Apr. 2017 - Jun. 2017.
This project studies lotus effect and spinodal decomposition phenomenon using finite difference method (FDM).
Underground Explosion
Concurrency Course Project at Ecole Polytechnique, Apr. 2017 - May 2017.
The propagation of the acoustic waves engendered by an explosion of a bomb is simulated using finite difference method. JAVA Threads are used to perform parallel computing.