Analytics India Magazine got in touch with Abhishek Bhandwaldar, Research Engineer at IBM to understand his machine learning journey. Abhishek has a Master’s in Computer Science from the University of North Carolina. “It is important to have a basic understanding of the different topics in the field to make sure you end up in the area you feel most passionate about,” says Abhishek.
AIM: What drew you to machine learning?
Abhishek: My introduction to AI was through video games. Then, I read about how ‘Deep Blue’ devised long-term strategies and beat an expert opponent in chess. This was quite fascinating because of its potential applications in solving real-world problems. It prompted me to work on reinforcement learning projects. We didn’t have a separate machine learning course, so I ended up taking a few Coursera courses on machine learning, probabilistic graphical models, and neural networks.
I decided to explore neural networks. I also had, around this time, started participating in ML competitions on Hacker earth (This was before Kaggle had become a thing). ML competitions are a great way to get hands-on experience working on real-world problems as ML algorithms can be fragile and require proper hyperparameter tuning.
AIM: Tell us about your machine learning journey so far
Abhishek: I spent a fair amount of my undergrad at SKNCOE college, Pune University, trying to decide if I wanted to make my career in AI or become a game designer. My interest in AI was far too strong, and I ended up applying for a Master’s in computer science. I got my Master’s from the University of North Carolina at Charlotte. During my Master’s, one of the best decisions I ever made was opting for doing a Master’s thesis. This helped me get first-hand experience in research.
I did an internship in a startup called Botsplash, where I worked with real-world data to train components of conversational agents. Later, I joined the AI Experience Lab in IBM research, where we created an interactive game that would communicate complex research ideas to the general public. The demo was accepted at CVPR 2018 and is now available to the public (here). Later, I moved to the MIT-IBM lab to continue work on 3D simulator and reinforcement learning. The 3D simulator, called ThreeDWorlds, was the result of a collaboration between our team at IBM, MIT and Stanford.
AIM: How did you overcome the initial challenges?
Abhishek: Growing up in India, AI wasn’t something that was widely taught and a rigorously researched topic. High school computer science education at that time was limited to teaching programming languages. I always loved those classes, and I was good at C programming. But I do wish they had included at least some AI topics. When we got our first internet connection, Wikipedia became my go-to page. So when I started my undergraduate studies, I made most of it by getting my hands on a lot of good material, attending guest lectures, and doing many projects with some amazing professors.
AIM: What does your typical day look like at IBM?
Abhishek: I work as a research software engineer in MIT-IBM Watson AI Lab. My responsibilities range from writing and maintaining codes for research projects, setting up experiments and publishing research papers. Often, research projects involve exploring the research idea by going over previously published material and trying out their codebases, and rigorously testing the ideas by conducting a number of experiments. Some of our work in the past has led to publications in top conferences and workshops, public release of AI demos, patent filing and release of AI challenges.
One is expected to maintain a balance between innovation and business results. I do this in our lab, taking leads on projects wherever necessary, planning project timelines, ensuring deadlines are met on time, identifying and publishing intellectual property through patents, and volunteering at various IBM hosted events & University programs.
AIM: What can India do to bolster its research ecosystem?
Abhishek: Most US universities have strong funded research programs. Most of the funding comes from government and industry grants and fellowships, and donations from alumni and philanthropists. These research groups are a great place to gain research experience as early as during undergrad. This leads to higher chances of students pursuing a PhD degree and a career in research. The sheer number of colleges in India makes it difficult for the government or private companies to fund research groups. And without proper funding, it is hard to employ and maintain talents in labs.
Our current economy focuses more on increasing the number of jobs. This needs to change to create more diverse and impactful job opportunities that will encourage people to pursue different fields and promote personal growth throughout their careers. A good example of this can be seen in the United States, where this kind of thinking has produced some of the most innovative companies working in a wide range of fields (Uber, Tesla, Amazon, Blue Origin).
AIM: Could you tell us a bit about the research works you were part of?
Abhishek: Most recently, I have been involved in the Machine common sense group, which aims to propose a machine learning model that can exhibit the same common sense as a young child. For machines to successfully be able to have social interaction like humans, they need to develop the ability to understand the hidden mental states of humans. Our work aims to bridge this gap by proposing a dataset that probes core psychological reasoning concepts.
Our dataset is a collection of videos similar to the developmental studies but generated at a much larger scale with visual differences. We have also proposed two different ML approaches to solve the dataset. Firstly, we show a familiarisation video to the ML model where the agent behaves in an expected manner. Then, the model is given a test video where the agent can behave in either an expected or unexpected way. The model outputs a surprise score. We expect the model to be surprised when the agent behaves unnaturally or unexpectedly.
We also conducted experiments on adults where they are shown similar videos and indicated their surprise on a scale of 1 to 5. This human score is then compared with our machine learning model score. We found that our proposed model scores were very close to how humans rated the agent behaviour. Our work was recently accepted at ICML 2021 and can be found here.
AIM: What is your advice to students looking to pursue AI/ML?
Abhishek: There are a lot of resources available today for getting started. These include both introductory and advanced material. The best way to stay updated with new advances is to read research papers regularly. Moreover,
- One should begin by building a foundation first. This includes doing courses on calculus, probability, and statistics, and introductory AI.
- Machine learning is a broad field that overlaps with other fields like natural language processing, computer vision, robotics, and computational neuroscience. It is important to have a basic understanding of the different topics in the field to make sure you end up in the area you feel most passionate about.
- This can be done by going over introductory books and videos as well as blog posts by seasoned researchers. If you have taken an introductory course in ML, then you could start working on Kaggle problems. The aim should not always be to win a cash prize but rather to learn as much as you can from the amazing community that the platform has built.
AIM: What are your books/resources recommendations for ML enthusiasts?
Abhishek: One of the first books I read was ‘Artificial Intelligence: A Modern Approach ‘by Stuart Russel and Peter Norvig’. My first best introduction to machine learning was through the Coursera course on Machine learning by Andrew Ng and Deep neural network for machine learning by Geoffrey Hinton. Other resources include:
- “Reinforcement Learning: An Introduction by Richard Sutton and Andrew Barto.”
- The Deep RL course by UC Berkeley taught by Sergey Levine complements the book.
- For deep-diving into Bayesian learning and probabilistic graphical models, the Probabilistic graphical model’s course by Daphne Koller is the best (It’s available for free on Youtube).
- “Learning from the data” by Yaser Abu-Mostafa dedicating more time to concepts like regularization, bias-variance tradeoff etc.