Me
Me

Welcome

Bio

I am currently a Masters of Science in Computer Vision student at the Robotics Institute at Carnegie Mellon University. I perform research in computer vision and am focused on research that has large practical impact and that opens up new avenues of research. I am also the founder and president of The Computer Vision Club @ CMU, the first computer vision club in USA. Prior to joining CMU, I did a dual Computer Science and Mathematics degree at The University of Texas at Austin. I also started a company TexteDB where I created an in-memory database for text analytics.




Publications

Architecture Compression
Overview of Architecture Compression
Overview of Architecture Compression
Anubhav Ashok
Under review for ICLR 2019
Paper


Knowledge Factorization
Task of Knowledge Factorization
Task of Knowledge Factorization
Anubhav Ashok
Under review for CVPR 2019



N2N Learning: Network to Network Compression via Policy Gradient Reinforcement Learning
Overview of N2N Reinforcement Learning pipeline
Overview of N2N Reinforcement Learning pipeline
Anubhav Ashok, Nicholas Rhinehart, Fares Beainy, Kris Kitani
Accepted in ICLR 2018 (poster)
Code
Paper


Spatiotemporal Attention Maps for Action Recognition
Anubhav Ashok Code



Research Projects

Scene2Synth: Inserting synthetic objects into a dynamic scenes using lightfield videos
We develop a system that allows the user to photorealistically insert a synthetic object into a lightfield video, where the synthetic object can interact with the moving objects in the video realistically
The sphere on the right is a fake object while the real painter is reconstructed from the original video
The sphere on the right is a fake object while the real painter is reconstructed from the original video
Poster
Video


Perpetual Generative Adversarial Fooling of Neural Networks
An approach to using GANs to generate infinitely many, visually dissimilar samples to fool neural networks perpetually
We generate both images of dogs and its corresponding perturbation using GANs
We generate both images of dogs and its corresponding perturbation using GANs
Report


Better Stochastic Gradient Descent
Created a novel variant of stochastic gradient descent that consistently outperformed all current gradient descent variants (SGD, Adam, RMSProp, AdaGrad etc.) tested on the MNIST and CIFAR-10 datasets
Convergence of our method compared to others
Convergence of our method compared to others
Report


Improving Video Segmentation Using Region Proposals and FCNs
Introduced a novel approach to video segmentation inspired by two modern segmentation approaches
The images of the left are the baselines using a single approach, while the right image shows the result of our combined approach
The images of the left are the baselines using a single approach, while the right image shows the result of our combined approach
Poster
Original video
Semantic Segmentation
Object Proposals
Our approach


Spaghetti sauce generation using Machine Learning
This project was done under the Freshman Research Initiative at UT Austin.
For this project, I created a program to generate novel spaghetti sauces based on a database of sauces containing various flavor information. The final sauce produced by the program was prepared and presented to the class during the final presentation.
A picture of what the spaghetti sauce might have looked like
A picture of what the spaghetti sauce might have looked like
Scalable drug categorization
Implemented a scalable, parallelized clustering algorithm to cluster drugs by their effects on cells. Won Bronze Prize in HPCQuest 2011.
Dorky image of me holding the award
Dorky image of me holding the award

Presentation