How Andrew Ng’s course helped launch Akash Palrecha’s AI career

Akash Palrecha always had a propensity for physical sciences, and machine learning came naturally to him. He completed his M.Sc in Mathematics at BITS Pilani and landed a job at Pixxel, one of India’s leading space tech startups, while still at college. Recently, Pixxel launched its first fully-fledged commercial satellite ‘Shakuntala’- a high res hyperspectral commercial camera– with SpaceX’s Falcon-9 rocket.  

Analytics India Magazine got in touch with Akash to understand his AI journey.

AIM: What drew you to AI? 

Akash Palrecha: After the first year, students have two-three months of free time. I was looking at different fields of science, anything related to computers, where I could get my hands dirty and find what interests me. I started with development and coding and moved on to machine learning. My search ended when I took Andrew NG’s introduction to machine learning course. It made me very curious, and I could see a lot of possibilities. 

AIM: How important is it for AI/ML aspirants to start early?

Download our Mobile App

Akash Palrecha: In my case, a degree didn’t hold much weight because I have been involved with Pixxel from early on. After my first year, I turned to NG’s course. I went on to Kaggle like everyone: I tried finding some data sets, and I thought I’d make some models and see where it took me, but I couldn’t do it. I realised Andrew NG’s course required a lot of understanding. And there was a good amount of theory, but not much about how to do things practically.  That’s when I stumbled upon Fast.ai. Their course was focused on the practical aspects. Around my third semester, a friend told me about Pixxel. The startup was looking for people to build AI models for satellite imagery. I spoke to Kshitij Khandelwal (CTO) and got myself an internship. The bar was pretty low to enter Pixxel at the time. Even though I started early, I wouldn’t say it is necessary. If you go on Twitter, you will see all sorts of people in all sorts of age groups starting their AI journeys. Some people start after completing their PhD; you see a lot of physics PhDs getting into AI; I have also seen some high school students getting into AI. 

Unlike other physical sciences, to get started in AI, all you need is knowledge of some linear algebra and basic Python. However, getting into big tech companies like Google, Netflix, etc., require at least a master’s degree and preferably a PhD. If you want to get into world-class teams, you will have to start early in college, ideally before your second year ends. After that, you can start preparing for a good master’s program. You have to publish good papers to get into colleges like Stanford, Berkeley, CMU, and MIT.

AIM: How equipped is India’s education system to foster our AI ecosystem?

Akash Palrecha: I will first start with my experience at BITS Pilani. In the four years I was there (from 2017 to 2021), AI underwent a massive shift. In the initial two years, the courses were obsolete and had not been updated. However, many courses were updated to cover broad areas by the time I graduated.

When I spoke with friends from other colleges in Pune, Bangalore, etc., I found that the courses are very basic and don’t go beyond linear & logistic regressors. Another major problem is that the teachers themselves are not knowledgeable. I know it’s not good to make a claim like that. But it’s just been my personal experience and based on the papers that come out of Indian colleges. Only a handful of places like IIT Delhi and IIT Mumbai have good labs and come out with good papers. The colleges that come up with good papers have good courses.

AIM: You became a FastAI international fellow last year. How did that come about? 

Akash Palrecha: It’s unlike most fellowships in that there is no application for it. I got it because I established my presence in the Fast.ai forums. I used to be involved in discussions about how certain parts of the library can be changed to be more user-friendly, allow certain architectures, etc. One day, I got an email from Jeremy Howard, founder of Fast.ai, stating that I had been selected as the first-year international fellow. The course started in May 2020 at the University of San Francisco, and I had to attend it remotely. To get a fellowship at Fast.ai, you have to add value to the forums, and then people will upvote your contributions, etc. I have learned a lot from this fellowship. I got to collaborate and share knowledge with many smart and interesting people in AI. 

AIM: Can you elaborate on your work on hyperspectral reconstruction?

Akash Palrecha: Normal images will have three channels: red, blue, green, and black and white images on one channel. Then, you have multispectral images, with about eight to 10 channels. In multispectral, other than the red, green, and blue parts of the visible spectrum, you’re covering infrared and ultraviolet.

In hyperspectral, you have 150 to 400 bands. Around 100 bands cover parts of the visible and invisible spectrum in a lot of detail. So every pixel in the image has much more information. The project I did was part of a workshop in CVPR, one of the top conferences for computer vision. You are given an RGB image, and your model should be able to extrapolate it into a hyperspectral image. So it’s simply a three-channel to 150 channel mapping. 

We did not have the best performance in the competition because, unlike the other teams from Zurich, etc., we started barely two weeks before the deadline. Still, our model was the fastest model by a few multiples of the next fastest model. 

AIM: How does Pixxel use AI in its products?

Akash Palrecha: Satellite imagery is used for a lot of analytical purposes. Hyperspectral imagery provides more detail to even a single point on the ground. The technology can be used to detect methane gas leaks, oil leaks in the oceans, etc. You can also analyse the chlorophyll and moisture content in plants. We are building a lot of models involving hyperspectral imaging which will soon be offered on our platform. The clients can come in, capture imagery of a certain area and ask for an analysis report. As far as the models are concerned, there is no constraint. We go from the basic decision trees and linear regressions to advanced computer vision models like Mask R-CNN or transformer-based models. Each product is unique and is tailored to the needs of the user.

AIM: What type of datasets are used to train such models?

Akash Palrecha: Unfortunately, not a lot of open-source data is available for hyperspectral imaging because there aren’t any public satellites to get hyperspectral data in the first place. Data annotation is a difficult task too. We get our own data label, in conjunction with many research scientists, etc. But there’s a lot of open-source data for simple problems like detecting buildings and structures and roads, aeroplanes, etc. Pixxel uses imagery from satellites, like Sentinel, Spot-6, worldview, etc. We also do a lot of fusion imagery from certain source labels to generate our own. We have model-assisted labelling, and at the same time, we feed new labels with assistance from our existing models. 

AIM: What are the interesting projects you have worked on? What are your long term goals?

Akash Palrecha: Lately, I have become interested in efficient machine learning. For example, mobile net architecture: It allows companies to deploy computer vision on edge devices, phones, tablets etc. And YOLO is another object detection architecture that caught my eye. I recently did some work at Aalto University with two friends. The work involved extreme machine learning where your models have to classify inputs ( in millions). It presents a lot of challenges in terms of scaling during training itself. We built model orders of magnitude smaller than the previous state of the art approaches and almost twice as fast in inference. 

I have also worked on a paper during my stint at Harvard that got published in CVPR a few days ago. I, alongside Siddhant Kharbanda and Atmadeep Banerjee, worked on tracking objects throughout videos. We were able to create a model that is memory efficient, and yet works on much longer videos than other methods. 

A few years down the line, I see myself being involved in the grassroots levels of frameworks like Pytorch and TensorFlow and contributing to their codebase. I want to contribute to technology, down to the last bit of the computation and less along the lines of creating new models. I think that’s important because those contributions expand to everyone else in the community. It powers the whole research space across academia and industry. But to be honest, it’s much less about contributing to the community and more about my curiosity. Contribution is a consequence of that.