Ashwini Pokle

ashwinipokle [at] stanford [dot] edu

My Photo

I am a second year Master’s student in Department of Computer Science at Stanford University. I have been working as a Research Assistant at Stanford Vision and Learning Lab (SVL) under supervision of Prof. Silvio Savarese since November 2017. I devote most of my time to the JackRabbot project. Prior to joining Stanford, I worked as a Software Development Engineer at Amazon for 2 years where I was a part of the Prime Video team and Fulfillment Center Technologies team.

I received my Bachelor’s in Computer Science from Birla Institute of Technology and Science, Pilani, India, where I worked on writing parallel and distributed versions of dimensionality reduction algorithms. I have spent a summer at USC Melady lab working under guidance of Prof. Yan Liu on studying techniques to improve convergence of stochastic gradient descent algorithm on heavy-tailed distributions. I have also interned at Bhabha Atomic Research Center, Mumbai in 2013.

My research interests include Computer Vision and Reinforcement Learning. Specifically, I am interested in understanding techniques to make AI more generalizable across multiple distinct tasks. I am also interested in other related research directions like improving sample efficiency of learning algorithms, overcoming the problem of catastrophic forgetting in AI, learning efficiently from demonstrations etc.


X. Zang*, A. Pokle*, M. Vazquez, K. Chen, J. C. Niebles, A. Soto, and S. Savarese. Translating Navigation Instructions in Natural Language to a High-Level Plan for Behavioral Robot Navigation. In Proceedings of Conference on Empirical Methods in Natural Language Processing, 2018.
[ website ][ code ][ paper ]

A. Pokle, R. Martın-Martın, P. Goebel, V. Chow, H. M. Ewald, J. Yang, Z. Wang, A. Sadeghian, D. Sadigh, S. Savarese, M. Vazquez. Deep Local Trajectory Planning and Control for Robot Navigation. In Proceedings of IEEE International Conference on Robotics and Automation, 2019


Defendr: A robust model for Image Classification
CS 231n Course Project with Apoorva Dornadula and Akhila Yerukola, Stanford University
We implemented an image classification model that is robust to black-box adversarial attacks like FGSM and PGD. The model uses a DUNet to denoise adversarial images. It also uses adversarial logit pairing in the objective to train a robust classifier. The model was able to achieve classification accuracy of 86.6% on tiny ImageNet (12,000 data points, 300 classes). This work was selected as the second best poster.

Visual Question Answering
CS 224n Course Project with Stefanie Anna, Stanford University
We implemented a visual question answering system that can provide single-word answer for a wide-variety of questions about an image. The model used Stacked Attention Network and was able to achieve accuracy of 47.11% on VQA v2.

Automated Question Answering
CS 221 Course Project with Amita Kamath and Akhila Yerukola, Stanford University
Suggested modifications to Dynamic Memory Network (DMN+) that improved convergence rate (upto 20% on 18 QA tasks) and improved accuracy (2% - 6% on 4 QA tasks) on bAbI-10k dataset.

Deep Reinforcement Learning for long term strategy games
CS 229 Course Project with Akhila Yerukola and Megha Jhunjhunwala, Stanford University
We implemented a hierarchical DQN on Atari Montezuma’s Revenge and compared the performance with other algorithms like DQN, A3C and A3C-CTS. The agent learnt a strategy to cross the first room and get a consistent reward.
[poster] [report]

Analysis of Emergent Behavior in Multi Agent Environments using Deep RL
CS 234 Course Project with Stefanie Anna, Stanford University
Implemented parameter-sharing DQN, DDQN and DRQN for multi-agent environments and analysed the evolution of complex group behaviors on multi-agent environments like Battle, Pursuit and Gathering. We observed emergence of interesting group behaviors in these games. In Battle, the particles learnt to cooperate to success- fully capture and kill the particles of the opponent team. In Pursuit, we observed that the predators learned to cooperate and form enclosures to trap the prey.
[poster] [report]

Parallelizing Dimensionality Reduction Algorithms
work done with Prof. Navneet Goyal, BITS Pilani, India
I worked on designing and implementing multi-core, multi-node and hybrid (multi-node + multi-core) implementations of PCA and SVD for dense matrices using algorithm for parallel TSQR factorization proposed by A. R. Benson et. al. in “Direct QR factorizations for tall-and-skinny matrices in MapReduce architectures.” I was able to achieve an average speedup of 4X in SVD and 8X in PCA for dense matrices of size up to 100 billion rows X 48 columns.

Robust Stochastic Gradient Descent
work done with Prof. Yan Liu, Melady Lab, University of Southern California
I worked on designing and implementing an algorithm for robust Stochastic Gradient Descent that aimed at improving the convergence rate of SGD on datasets with large variance and on heavy taled distributions. We were able to achieve upto 1.5x speedup on a randomly generated synthetic dataset of size 1000 with 50 random outliers.

Library for Online Adaptive Subgradient Methods
work done with Prof. Yan Liu, Melady Lab, University of Southern California
I implemented a C++ library for adaptive subgradient methods with primal-dual gradient update and composite mirror descent update for L1, L1-ball projection, L2 and L∞ regularization. These online algorithms were based on work by John Duchi et. al. in “Adaptive Subgradient Methods for Online Learning and Stochastic Optimization”.

Personalizing search results on the basis of user's Twitter activity
CS F469 Course Project, BITS Pilani, India
Performed concept tagging on the content of tweets and re-tweets to obtain concepts associated with a user. Performed taxonomical classification of these concepts into ∼ 30 broad classes and generated user-interest profiles. Re-ranked the search-results returned by search-engines on the basis of user's interest profile. Peri- odically updated the user profiles to account for variation in interests.

Implementation of Parallel Page Rank algorithm
CS F422 Course Project, BITS Pilani, India
Implemented multi-node, multi-core and hybrid (multi-node + multi-core) implementations of per- sonalized PageRank algorithm. Validated correctness of algorithm for datasets with spider-traps, disconnected components and dead-ends.

Implementation of compiler for a custom language
CS F363 Course Project, BITS Pilani, India
Implemented a compiler for a custom Matlab-like language. This was a team project and I focussed on implementing lexer, parser, abstract syntax tree and code generator.


Nutanix Women in Tech Scholarship, 2018
Regional Winners (India - Europe region), IC Tech Winter Hackathon, Amazon, 2017
Winner, AFT India Hackathon, Amazon, 2016
Merit Scholarship, BITS Pilani, 2011-2015
GE Foundation Scholar-Leader, GE Foundation, 2013