Karan Taneja

Karan Taneja

Researcher and student. Creating AI tools for education at Georgia Tech.

Blog posts

2020

How to Undergrad

20 minute read

Published:

As I end my undergraduate journey, it’s hard to not feel nostalgic, especially amid this sad pandemic because of which this journey was shortened by nearly 3.5 months. It was, in many ways, the best time of my life but, most importantly, I developed a better understanding of this world and of myself. A lot of these lessons helped me navigate through my undergraduate years and I wish to offer some suggestions to my juniors who are starting out and to those who are curious about how they can do better.

2019

Batch Arm Pulls for Stochastic Multi-Armed Bandits

7 minute read

Published:

I took the course Advanced Concentration Inequalities in Autumn 2019 offered by Prof. Sharayu Moharir at IIT Bombay. This is a blog post about the work I did for the course project with Shashwat Shukla and Sucheta Ravikanti on designing regret minimization algorithms for the problem of batch pulls in stochastic multi-armed bandits.

Voice Conversion using GANs: An Extensive Review

18 minute read

Published:

For my project on code-mixed speech recognition with Prof. Preethi Jyothi, I did a literature review of voice conversion and found a lot of recent papers that used GANs for the problem. So I decided to write this post to summarize my review. I’ve also attached the slides and references I used in my group meeting at the end of this post.

Compressed Sensing and Dictionary Learning for Video Capturing

12 minute read

Published:

I took the course Recent Topics is Analytical Signal Processing in Spring 2019 offered by Prof. Animesh Kumar at IIT Bombay. This is a blog post about the work I did for the course project with Anmol Kagrecha and Pranav Kulkarni on using compressed sensing and dictionary learning to alleviate tradeoff between temporal and spatial resolution in videos based on work by Hitomi et al. 2011 and Liu et al. 2014 and extending it to use different sampling strategies that can improve the performance but are not necessarily constrained by hardware implementations.

Semantic Image Inpainting for Medical Images with Deep Generative Models

14 minute read

Published:

I took the course Medical Image Computing in Spring 2019 offered by Prof. Suyash Awate at IIT Bombay. This is a blog post about the work I did for the course project with Pranav Kulkarni on using GANs for image inpainting based on work by R. Yeh et al. in CVPR 2017 and extending it to use VAE for the same problem and showing that VAEs perform better than GANs for our task.

2018

Computer Graphics: Music Box Modelling, Rendering and Animation

8 minute read

Published:

I took the course Computer Graphics in Autumn 2018 offered by Prof. Parag Chaudhuri at IIT Bombay. This is a blog post about the work I did for the course project with Pranav Kulkarni on modeling a music box with a humanoid figure and a giraffe in a realistic room with lighting and texture and rendering the animation of dancing characters with camera movement on a user-specified Bezier curve. The project was implemented in OpenGL. An example of animation video generated from our interface can be be seen on YouTube.

Tutorial on Attention-based Models (Part 2)

19 minute read

Published:

In part one of this series, I introduced the fundamentals of sequence-to-sequence models and attention-based models. I briefly mentioned two sequence-to-sequence models that don't use attention and then introduced soft-alignment based models. In this post, I’m going to discuss about various monotonic attention mechanisms.

Tutorial on Attention-based Models (Part 1)

37 minute read

Published:

Teaser: The task of learning sequential input-output relations is fundamental to machine learning and is especially of great interest when the input and output sequences have different lengths. Applications such as speech recognition, machine translation, document summarization, image captioning and many more can be posed in this format. In this post, we explore one of the popular ways to solve such problems using 'Attention Based Models'.