Misc.
Projects
 | Improving Semantic Relation Prediction using Global Graph Properties Course project for 10-708: Probablistic Graphical Models We extended ERGMs graph motifs to include combinatorial features to constrain local predictions and replaced negative sampling with hierarchical softmax, improving computation efficiency and performance on infrequent words. |
 | Few-Shot Learning for Images Course project for 11-785: Intro to Deep Learning We proposed a weighted distribution calibration to alleviate bias of distribution of novel classes by generating data from a calibrated distribution computed using transferred statistics of all base classes. |
 | Action Segmentation on Large Egocentric Cooking Dataset Course project for 11-777: Multimodal Machine Learning We proposed the first multimodal approach in video action segmentation that refines unimodal segmentation results with a video-to-text retrieval module. |
 | Emotion Detection with Ensemble-based CNN Final project for 10-701: Intro to Machine Learning We obtained large performance gain applying transfer learning and ensemble modeling and improved interpretability using occlusion-based saliency maps. |
 | Edge Term project for 15-112: Fundamentals of Programming and Computer Science Inspired by Monument Valley, I created an algorithm to generate random solvable mazes in a simulated 3D space using Blender and Panda3D. |