Evaluating Multisense Word Embeddings (CS 238)
We investigate the problem of multiple word senses in different contexts. We compare 'traditional' word embeddings like word2vec/Glove with generative probabilistic 'multisense' ones.
Image Compression using Deep Learning (CS 231N)
Benchmarked Conv. Autoencoder and Conv. VAE with RNNs and GANs, proposed a better GAN loss function specific to image compression
Exposing Racism and Sexism using Deep Learning (CS 224N)
Built DL model to classify tweets and achieved best F1 performance compared to previous end to end neural models reported in literature.
Detecting Insults in Social Commentary (CS 229)
Comparative study and benchmarking of BOW models with DL models on the largest dataset of user annotated Wikipedia Comments.
Community Detection in Graph using Semantic understanding of Text (CS 224W)
Proposed a novel method of detecting communities using a mixture of modularity maximization, sentiment minimization, spectral analysis and node embeddings.
Quantifying Pathway Heterogeneity in Protein Folding
Built Directed graphical models (networks) to compute the number of pathways over an exponentially large state space of a protein. These models are sampled using MCMC (Metropolis-Hasting) methods and clustered using MCL based on random walks.
- Extracting the Hidden Distributions Underlying the Mean Transition State Structures in Protein Folding: J. Phys. Chem. Lett., 9 (7), 2018
- Toward a quantitative description of microscopic pathway heterogeneity in protein folding: Phys Chem Chem Phys, 19, 2017
Generative Pixel Recurrent Neural Network
(Independent project wirh Prof. Kaushik Mitra)
Conceptualized a deep convolutional recurrent neural network to generate 4D lightfield images by modeling long range spatial dependencies.