Proposed a novel approach to introduce reconstructive feedback in the ALIGAN/BIGAN framework by modifying the discriminator's objective to correctly identify more than just two joint distributions of image-latent vector pairs, performed quantitative and qualitative evaluations over baselines on SVHN and CelebA datasets.
Undergraduate Project under Prof. Piyush Rai and Prof. Vinay Namboodiri. Designed a novel modelling technique for joint image and video generation that simultaneously learns to map latent variables with a fixed prior onto real images and interpolate over images to generate videos. Publication was accepted in WACV 2020.
Research project under Prof. Piyush Rai. Studied various approaches for learning disentangled representations of sequential data such as using using new adversarial loss terms, factorized hierarchical priors and exploiting the probabilistic model and architecture of the LSTM based autoencoder to promote disentanglement, implemented a Variational Autoencoder model for disentangling of time invariant content and dynamics in sequential data (Mandt et al.) using Pytorch and experimented with modifications in the probabilistic model.
Implemented Deep Q-Learning and Policy Gradient methods for Atari Games using PyTorch and OpenAI Gym along with various classical RL methods using Numpy such as Dynamic Programming (Policy and Value iteration), Monte Carlo (Epsilon-greedy and off-policy), TD Learning (Q-Learning and SARSA) and Q-Learning with Function Approximation. Presently studying state of the art variants of actor-critic methods.
Project under Programming Club, IIT Kanpur. Studied various encoder-decoder based architectures for image captioning and implemented the model described in Show, Attend and Tell (Xu et al.2015) using Tensorflow. Used MS COCO dataset for training and evaluation.
Research project under Prof. Nisheeth Srivastava. The aim is to study the effect of variation of multiple physical parameters in animated situations to determine the underlying cause of inference of emotion and social situations without language.