To see a complete list of my projects, see my CV. Below, is a list of some of my selected projects.
  • Proposed a spatio temporal end to end module that effectively exploits temporal redundancies while maintaining real-time speed for video colorization task
  • Spatial Transformer Network was used to encode optical flow features for motion estimation and motion compensation to reduce flickering noise in the output.
  • Early fusion, slow fusion and 3D convolutions were explored for the joint processing of multiple consecutive video frames.
  • This report proposes incorporation of attention and memory based components in basic architecture of Visual Dialog to reduce reliance of the embedding of an entire history.
  • The role of choice of memory block in RNN and word-vector is explored for improving the performance of the model.
  • [Report]
  • Implemented Siamese Recurrent Neural Network to measure semantic similarity between two sentences.
  • Used Quora 0.4 million sentence pair dataset , 100 dimensions GloVe word-vector as input to the network and got 0.74 accuracy on test dataset.
. View Source
  • Trained an agent using Asynchronous Advantage Actor-Critic (A3C) to learn task oriented language grounding by performing specific actions in a Doom game, through natural language commands.
  • Proposed an end to end neural architecture to efficiently encode the visual temporal reception of agent’s 3D environment.
  • VizDoom API along with SLADE editor used to construct different environment for an agent.