May 2022 – Sep 2024 · 2 yrs 4 mos · Hyderabad, Telangana, India

• Training Deep Nets on Speech Data

May 2021 – Jul 2021 · 2 mos · Hyderabad, Telangana, India

• The Speech Recognition group at IDC (PPO Received)
• Worked on Audio-Visual ASR to improve recognition performance in noisy environments by incorporating visual-features into the ASR pipeline.
• Developed a proof-of-concept that gave a 6% relative improvement in Word-Error-Rate(WER) compared to the baseline that did not use visual features, under noisy environments.

Apr 2020 – Apr 2021 · 1 yr · Los Angeles, California, United States

• Developed algorithms for the structure learning of Probabilistic Circuits , a model class allowing for tractable inference of complex probabilistic queries.
• Derived a theoretical connection between the independences induced by a structure and its log-likelihood to provide a new perspective for structure evaluation.
• Developed a novel algorithm to learn stronger ensembles of such circuits.
• Implemented the framework in Julia which gave state-of-the-art results on 14/20 density estimation benchmarks.
• Work published at TPM Workshop, UAI'21.
• Research Areas : Probabilistic Graphical Models, Probabilistic Inference, Generative Modelling, Learning Structure from Data

Apr 2020 – Jul 2020 · 3 mos · Linköping, Ostergotland County, Sweden

• Worked on the problem of Graph Representation Learning with the downstream task of link-prediction
• Experimented with various variants of GraphSAGE, Graph-Attention-Networks, GraphGANs and Gated-Attention-Networks
• Formulated and implemented a DFS-based aggregation scheme to capture community structures in graphs

May 2019 – Sep 2019 · 4 mos

• Worked under The Julia Language to add GANs, RL models and Neural Image captioning networks to the model zoo of Julia.
• Added models like pix2pix, cycleGAN, SRGAN, Neural Image Captioning.
• Created a library along the lines of the openai-baselines in pure Julia and added PPO, TRPO and Deep Recurrent Q Learning to it.
• Trained the reinforcement learning models on control-problems on the Open-AI-Gym environments.

Mar 2018 – May 2021 · 3 yrs 2 mos · Kharagpur Area, India

• We are a group of robotics enthusiast researchers and developing novel solutions in the realm of autonomous driving.
• We are working towards converting Mahindra e2o into a fully operational self driving car.
• My work is mainly as a part of the planning and perception team wherein I research on developing robust algorithms that work under unstructured environments. This amounts to testing them on simulation and deploying them further on the actual car and a husky robot.
• Created and trained a Mobile-Net-SSD object detection pipeline for real-time detection of traffic-signs. Obtained 50+FPS on a 970mx GPU.
• Implemented various adversarial and non-adversarial approaches for real-time detection of lanes.
• Worked on real-time monocular-road-segmentation for identifying drivable region in the unstructured roads of the University Campus. Obtained real-time speeds of 30+FPS on a 970mx GPU.
• Worked on the Frenet Planner for jerk-minimising highway trajectory planning. Implemented the pure-pursuit path-tracking strategy and tested the implementation on a real car.
• Experimented with various deep-reinforcement-learning algorithms for training agents for autonomous indoor navigation.
• We were adjuged 2nd in the autonomous navigation challenge in the Intelligent Ground Vehicle Competition, 2018 and 2019.
• Worked on the planning and perception part of the software stack.
• Work resulted in three first-author publications in deep-learning and robotics.

Mar 2018 – Feb 2019 · 11 mos · Kharagpur Area, India

Mar 2018 – Aug 2018 · 5 mos · Kharagpur Area, India

• Research focussed on strategy and planning for robots playing robot soccer for the small sized league in robocup.
• Implemented a Fuzzy logic based multi-agent passing mechanism.
• Worked on path planning and implemented RRT,RRT* and RT-RRT* from scratch in C++.

Aug 2017 – Mar 2018 · 7 mos · Kharagpur, West Bengal, India