Boston University

Convolutional Neural Networks for Unsupervised Image Segmentation

Research advised by Professor Brian Kulis.

• Design an end-to-end Convolutional Neural Network Architecture for fully-unsupervised image segmentation without any labeling information.

Efficient Deep Generative Models for Unsupervised Representation Learning

Research advised by Professor Brian Kulis.

• Learn the underlying lower-dimensional representation for input image data on the hidden layer.

• Train a deep generative model for the purpose of unsupervised clustering task in the hidden space.

Age and Gender Prediction on Twitter Users

Research Assistant advised by Professor Margrit Betke, Image and Video Computing (IVC) Lab.

• Efficient Deep Generative Models for Unsupervised Learning.

• Age/ gender/ ethnicity prediction on the data of Twitter users profile images.

Harvard Medical School

Computational prediction of protein-DNA interactions based on sequences information

Graduate Research Fellow at Department of Systems Biology

• Develop a new computational method for predicting protein-DNA interactions.

• Implement large-scale scientific computing in parallel and distributed environments.

Harvard University

Intervention and Outcome Predictions in the ICU

Research advised by Professor Finale Doshi-Velez.

• Design a recurrent neural network (RNN) model to simulate multidimensional physiological time series of patients during vasopressor administration.

Batch Mode Active Learning and Its Application to Astronomy

M.E. Thesis advised by Professor Finale Doshi-Velez and Dr. Pavlos Protopapas.

• Developed a batch-mode cost-sensitive active learning approach that not only exploited uncertainty and representativeness of the whole unlabeled dataset but also took annotation cost into consideration.

• Designed a selection criterion that combined uncertainty and representativeness by usin