Last Updated: March 2021
Hey there! I’m Vikas Nataraja, a researcher at the Laboratory for Atmospheric and Space Physics (LASP). I design and experiment with various Machine Learning/AI algorithms and models that are designed to improve the state-of-the-art. I started this blog mainly because I love writing and blogging but also so I could learn basic web design. I’ll be blogging about anything that I find interesting - bias in AI algorithms, reinforcement learning, latest ML papers etc. I want to keep improving this website, so feel free to get in touch to tell me how I can or tell me what issues you want me to write about on this blog.
I graduated from the University of Colorado Boulder in May 2020 where I worked on a lot of different things ranging from robotics and self-driving cars to aerospace and atmospheric science. I focused on the perception side of things with camera and lidar data, designing ML and CV algorithms. More recently, I have been working on bias in AI and convolutional neural networks and some other cool stuff. Check out my blog where I write about ML topics.
December 2019 - present
Currently, I work as a Research Assistant at the Laboratory for Atmospheric and Space Physics with Dr. Sebastian Schmidt and his group to use machine learning to predict Cloud Optical Thickness from cloud radiances. The idea is to design a CNN to quickly estimate the Cloud Optical Thickness or COT based on a 3D radiative transfer model. In simpler words, calculate the optical thickness of clouds in images taken from a Sky-View camera by taking into account the radiative transfer (geometry and spectral distribution of radiation) through clouds. We started with a PSPNet-based framework/backbone but quickly switched to other complex architectures which have produced some incredible results. I’m pretty pumped about it, stay tuned for our paper!
My internship at an awesome startup in Pittsburgh
In the last couple of years, I have had the opportunity to work on some truly fascinating topics. In the summer of 2019, I interned as a Software Engineer with Allvision IO in Pittsburgh, PA where I worked on building and training a Faster R-CNN and an SSD model to detect license plates in a video stream. In my 3 months at Allvision, I wrote an end-to-end pipeline to grab the images from the dataset, preprocess them, feed them into my model which produces 2D coordinate bounding boxes which are then sent to OpenALPR for license plate character recognition. My model performed quite well on Pittsburgh’s Strip District and East Liberty datasets which were the only two available to me at the time.
During my time at CU Boulder, I was a Teaching Assistant for a couple of courses - ECEN 2703 Discrete Math for Computer Engineers with Dr. Fabio Somenzi and ATOC 4815 Data Visualization with Python with Dr. Sebastian Schmidt. While both courses were immensely exciting to teach, the latter was so much fun because I had the chance to create new course content and teach Python and Machine Learning to students who had never even heard or experimented with it. It presented a new challenge to me since I mostly run in Computer Science and AV circles where it is quite safe to assume that people have heard about these kind of topics and brought me back to my roots and forced me to re-learn some topics I had forgotten about. One of my more exciting ventures for sure!
Lead Software Engineer - ADCS at Maxwell CubeSat
September 2018 - May 2020
For the entirety of my grad school life, I worked for the ADCS Software Engineering Team for the Maxwell CubeSat mission and I held various roles including Volunteer, Software Engineer and Team Lead. I worked on Attitude Determination and Control Systems problems for the satellite. Specifically, I developed Flight Software to use perception (from sensors) and control the motion of the satellite using actuators. Currently, it is scheduled to be delivered to the US Air Force Research Lab by late 2020/early 2021 with a tentative launch date in the fall of 2021. The biggest reason why I joined this project in the first place in spite of it not being my focus area was the opportunity to take my work on Perception systems in AVs and use them in the real world and learn about control systems which hasn’t always particularly been my cup of tea. Plus, the chance to launch an actual satellite that would fly in space was hard to turn down and it was one of the best experiences I’ve ever had! Check out our paper here.