Worked on developing off policy Reinforcement Learning algorithms to successfully learn navigation policies in urban environments, with multiple dynamic agents, of CARLA simulator.

TA for 10601-Introduction to Machine Learning course taught by Prof. Matt Gormley.Responsibilities include conducting office hours, preparing assignments and grading.

I worked under the supervision of Prof. Vinay Namboodiri through an official collaboration. I worked in the following projects:1) Neural Network Architecture Compression(CNNs)∗ Formulated an algorithm based on robustness of filters to fatuous auxiliary loss functions.∗ Achieved an impressive reduction in FLOPS by 80% and 44.5% in VGG16 and ResNet-50,respectively, on ImageNet.∗ Demonstrated the flexibility of our approach by achieving 33% reduction in FLOPS in C3D on UCF101 and 36.8% reduction in MobileNetV2, which is already a compact network, on CIFAR100.∗ This work made it feasible to run image recognition and video action recognition networks on TV(Tizen platform) with no additional computational support. This work is published in WACV 2019 conference.2) Video Action detection and recognition:∗ Worked on both RNN and 3D CNN based approaches for action detection and recognition on ActivityNet dataset.∗ Implemented state of art works like R-C3D for temporal action detection and recognition.∗ To perform action recognition for live video streams, we worked on RNN based approaches where I proposed a variant of RNN which has non-chaotic strange attractors, in dynamical systems perspective.∗ Incorporated hard temporal attention module using stochastic binary neurons, for untrimmed videos, to select frames to be considered while recognizing action and trained end to end using straight through estimator.3) Few-shot learning (not in collaboration)∗ Designed a system to recognize STB box providers using logos from few images and account for variations over time.∗ Trained a siamese network using triplet loss function and hard negative mining and ported it to TV using tiny-dnn framework.

Worked on detecting the indoor location of user using bluetooth signal strengths from beacons deployed in indoor environment. To account for dynamic obstacles in indoor environment, we applied machine learning techniques to solve this task. We used time series filtering methods to remove fluctuations in Bluetooth signal strengths due to noise. Trained machine learning models like random forests, Ada-Boosting, SVM and Neural networks for coarse positioning into grids(calculated using indoor layout information).