Working on ML models, pipelines, and tools of the Motion Planning and Prediction team for 3+ years at Aurora and Uber ATG.

Checkout my Resume for more details.

Paper Link: https://arxiv.org/abs/2006.13916

Published in ICLR 2021, and was a spotlight talk in NeuRIPs 2020 and ICML 2020

We propose a simple, practical, and intuitive approach for domain adaptation in reinforcement learning, where we learn with small data from real world, leverage simulation data, to perform well in the real world.

We focus on generating faces corresponding to certain facial features using Generative Adversarial Network (GAN) to pursue our task. We intended to develop a GAN to assist police sketch artists. Here the descriptive facial features are the inputs( e.g. hair color, eye color, etc) to the GAN which generates a face corresponding to those features. We also try to control the individual facial features to tweak the changes in the face generated.

Github link: https://github.com/asawaswapnil/Deep-Learning-Project-Face-Generation-With-GAN

The most popular reported metrics namely Accuracy, Precision, Recall, and F1-score are not good indicators when the data is imbalanced. Also, some of these matrices are not symmetric i.e. if we exchange just the labels of positives and negatives, good measures may turn poor and vice versa.In this work, I did a robust comparison of various evaluation metrics on data imbalance and symmetry and analyse their results. Furthermore, I proposed a new evaluation metrics, TPNR, which is robust to these changes.

GitHub link: https://github.com/asawaswapnil/Machine-Learning-Final-Project-With-Python

We modified the ideas for Bidirectional Attentive Fusion for Dense Video Captioning using 2D convolutions with MobilenetV2 and LSTMs, achieving 8.4% reduction in training time and very similar accuracy as state-of-the-art.

GitHub Link: https://github.com/asawaswapnil/DenseVideoCaptioning

The basket(agent) doesn’t know anything about the game initially. The goal is to catch all the apples. Using reinforcement learning, it learns in 350 iterations and plays like an expert. If the basket catches the falling apple, the score increases. Otherwise, the score resets. The number of iteration/ episodes and the score can be seen at the top right corner.  (In the video you can see it learning initially by exploring options, and after 43 sec(from 350 to 500 iterations), it has learned everything).  Github link: https://github.com/asawaswapnil/intellegent-basket .

In order to help restaurants to keep better management in their place, we
try to predict an approximate range of visitors will be visiting the restaurant for future dates using multiple datasets from different sources.

GitHub Link: https://github.com/asawaswapnil/customerTrafficforRestaurant