Mar 2022 – Apr 2024 · 2 yrs 1 mo · Mumbai, Maharashtra

• UPSC CSE preparation

Apr 2021 – Feb 2022 · 10 mos

Nov 2018 – Mar 2021 · 2 yrs 4 mos

May 2018 – Aug 2018 · 3 mos · Greater Boston Area

• Learning suitable and well-performing dialogue behaviour in statistical spoken dialogue systems has been in the focus of research for many years. While most work which is based on reinforcement learning employs an objective measure like task success for modelling the reward signal, we proposed to use a reward based on user satisfaction. We evaluated different methods such as Linear Regression, SVM, RNN, etc. to compare the pros and cons of each method. The model will later be used to assist the training of other components of conversational AI such as dialogue manager, etc. to be more robust and converse in a more satisfactory manner.

Aug 2017 – Dec 2017 · 4 mos · Mumbai Area, India

• We frame question generation as a transduction problem starting from a Freebase fact,
• represented by a triple consisting of a subject, a relationship and an object, which is transduced
• into a question about the subject, where the object is the correct answer. Intuitively, one can
• think of the transduction task as a “lossy translation” from structured knowledge to human
• language, where certain aspects of the structured knowledge is intentionally left out. We are
• using seq2seq model in Tensorflow with custom designed GRU cells and attention.
• Planning to submit the work at NAACL 2018.

Jul 2017 – Dec 2017 · 5 mos · Mumbai Area, India

• Neural network-based methods represent the state-of-the-art in question generation from a text.Existing work focuses on generating only questions from a text without concerning itself with answer generation. We present a novel two-stage process to generate question-answer pairs from the text. For the first stage, we present alternatives for encoding the span of the potential answer in the sentence using pointer networks. In our second stage, we employ sequence to sequence models for question generation, enhanced with rich linguistic features, named-entity alignment to handle rare words. Finally, global attention and answer encoding are used for generating the most relevant question. Our experimental results validate the significant improvement in the quality of questions generated by our framework over the state-of-the-art. A research paper submitted at AAAI 2018. Demo of our system is accessible on https://goo.gl/2XpXPv.

May 2017 – Jul 2017 · 2 mos · Zürich Area, Switzerland

• IT equipment failure significantly contributes to data center down-times. We try to predict such equipment failures a week in advance. For each such device, we record the utilization monitoring data and event handler logs. While potentially containing a wealth of insights, the data is difficult to mine effectively, owing to varying length, irregular sampling and missing data. We devised a model based on Recurrent Neural Network with LSTM cells to predict device failures on such data. To combat the imbalance in our dataset, we employed data augmentation and weighted loss function. We experimented with bi-directional LSTMs and sequence labelling techniques to maximise the precision. Hyper-parameter tuning was performed using grid-search. Our model improved the F1 score by 24% over random forest baseline classifier.

May 2016 – Jul 2016 · 2 mos · Liège Area, Belgium

• Working under Prof. Louis Wehenkel and Prof. Damien Ernst in the area of Machine Learning mainly Deep Reinforcement Learning. Studied and implemented various concepts such as Universal Value Function Approximators, Monte Carlo tree search for planning and pruning etc on Pong and Tetris with deep neural networks to achieve Super AI performance. The Learning agents managed to defeat OpenAI's gym pong player in 80% of the total games played.
• Contributed to the DeeR(Deep Reinforcement) Learning Framework by working on Planning(look ahead) using Monte Carlo Tree Search techniques on the game tree which significantly decreased the complexity and while still managing to explore new trajectories(avoiding over fitting). Made the framework compatible with TensorFlow by adding the functionality of Q-Networks written in TensorFlow. Working on a research paper for the framework as a co-author.
• Also, worked on a theoretical problem of Using discount factor lower than evaluation discount factor while training on approximate MDPs. Currently working on a hypothesis relating to the difference of the value functions of the two discount factors.

Nov 2015 – Jan 2016 · 2 mos · Mumbai Area, India

• Worked on Information Retrieval scripts in Python to perform on well linked websites. Developed a text mining tool to extract relevant information from crawled data. Used the concepts of Vector spaced models and computed cosine scores to classify results.