I was in a team of 3 members, we were given a broad area “Privacy-Aware Machine Learning” and were asked to come up with a problem statement in that area. The problem statements which we came up with finally were:
- Incorporating personalization in existing differentially private recommender systems
- Differential privacy in sequential based recommender systems

For the first problem, we incorporated the concept of user and item-based privacy level in recommender systems. To deal with the second problem we explored Reinforcement Learning and LSTM based recommender systems and settled to apply differential privacy to LSTM based recommender system by updating a library for the same.

*Patent (Application No. 16041182) titled ’Recommendation System Based on Individualized Privacy Settings’ filed at the US Patent office.

Technical Stack: Python, Tensorflow, Theano

This dissertation deals with enhancing the images using Convolutional Neural Networks to get the lost information due to degradation during acquisition of image and make the image visually appealing and informative. This project was based on two contexts, one being removing artefacts due to the acquisition of an image and the other being improving image quality to make it more visually pleasing. Worked on two problem statements -
- Joint HDR and Depth Estimation from varying exposure stereo image pairs
- Enhancing Images to make the objects more recognizable by state of the art CNN’s. (Dataset Used - UG2 challenge dataset)

The former uses multiple frames to convert the image into an HDR(High Dynamic Range) quality and we formed a siamese CNN based model which will give HDR image from a set of multiple exposure images. The latter uses only a single frame and we formed a novel convolution neural network(CNN) based model which will improvise real-world images to get good classification accuracy on the state of the art object detection models. Extensive experiments were done in both contexts based on GAN’s, Resnet, VGG etc.

Technical Stack: Tensorflow, Pytorch, Lasagne, OpenCV

Took part in a Microsoft AI Challenge. The task of this challenge was to mark the most relevant passage among some candidate passages which contains the answer to the user query.

Used glove model to embed the query passages pairs. Explored various models based on LSTM, Siamese Networks and GRU. Ran a attention based GRU Network which gave the best accuracy of 0.54. Was among the top 50 teams and entered the final phase of the competition.

Technical Stack: Pytorch

 A study on some classes of minimum divergence estimators indexed by one or two tuning parameters was done. The divergence studied were Density power Divergence (DPD), Logarithmic Density power Divergence (LDPD), Bridge Divergence and B-exp Divergence (BED).

Estimators generated from minimizing these divergences are all M-estimators. Each of these estimators takes Maximum Likelihood or Minimum L2  estimator as a special case. In each of these estimators, robustness is achieved through some down-weighting of outliers. Their robustness and efficiency were explored through extensive simulations and real data analysis.