Interview with Ajay Prasadh Viswanathan, AI Resident
Ajay Prasadh Viswanathan, AI Resident
Tell us about yourself and how you came to join the Cambridge AI Residency Program.
My journey in AI began when a friend challenged me with a simple puzzle, five years ago. Imagine someone drawing a circle on paper and asking you whether a new point is inside or outside the circle. This is easy, as you can see the circle.
But what if, instead of drawing the circle, if they just gave you some points inside the circle and some points outside the circle, and then asked the same question again? That seems impossible, like magic. My first impression of machine learning was that—a combination of mathematics and computer science to do something magical.
I did my Bachelor’s degree in computer science and Master’s degree in artificial intelligence, specializing in machine learning. I briefly worked for a start-up based in Oxford, UK, building digital characters using language processing. I felt that being a part of a larger research organization would help me learn more, grow faster, and build better machine learning systems. The AI residency program here at Microsoft Research Cambridge offered all of that. So here I am learning, applying, and playing with some of that magic.
Can you give us a high-level overview of what your first project was?
Our first project was with the HoloLens team and AI Resident partner Sebastian Dziadzio. The task was to generate a novel view of a given scene, when given the RGB and depth input from a Kinect. There were three key moments in the project that I fondly remember. Building a classic vision pipeline to transform from one view to another was really hard as we did not have prior experience in building 3D vision systems, and 3D vision textbooks are incredibly dry. So we got very excited when we got this classic (without machine learning) vision system to work. We built a machine learning pipeline on top of this classic vision system and we found that this worked better than a pure machine learning system. The final breakthrough moment was doing a live demo during TechFest, where our model worked pretty well for completely novel viewpoints, lighting, and people, all in real time!
What excites you the most about the field of AI and ML?
The same core set of algorithms are making state-of-the-art advances in many fields such as speech recognition, language understanding, and computer vision. Deep learning is having an impact on many production systems around the world, from self-driving cars to setting default parameters for compilers. So far, most of what we’ve seen is supervised learning, where computers learn to do things by seeing lots and lots of examples, which is different than how humans learn. Unsupervised learning and semi-supervised learning might guide us to build machines that are smarter. In many ways, I believe the best is yet to come.
What has surprised you the most about working and growing at Microsoft?
The breadth of the areas worked on at Microsoft Research Cambridge has surprised me the most. In one building, people are working on the HoloLens, storing data on glass, predicting RNA reaction rates, understanding computer code, confidential computing, human computer interactions, and machine learning. This creates a rich and diverse environment where exceptional people solve exceptional problems through ambitious research projects. Being around these people, understanding some of their work, and attending guest lectures has significantly enhanced my knowledge of what is possible to do with science and technology.
What advice do you have for the incoming AI Residents?
These are great pieces of advice I received from my research supervisor, Andrew Fitzgibbon.
Focus on and know the basics (programming, linear algebra, probability, statistics, basic calculus) really well. I usually fail to understand a complex system or paper when I struggle to understand one of the basic ideas.
At the same time, it’s important to, as Andrew says, “focus on everything.” Focusing on every aspect of what you are building is vital for making a real impact. These aspects could be as specific as coding and as general as design choices, but they all contribute to the success of a project.
Of course, you also have to make sure you’re making that real impact, such as someone reusing code or an algorithm you have written, someone referencing a result you have built or observed, or someone buying a product that you have built. Someone has to get real value from using what you have built, so keep an eye on that value from the beginning.
What books do you recommend for an aspiring AI/ML enthusiast to read to learn and grow?
The Deep Learning textbook by Ian Goodfellow, Yoshua Bengio, and Aaron Courville is very good. David MacKay’s book, Information Theory, Inference and Learning Algorithms, is great as well. It comes with a set of online lectures that are available on YouTube. I should mention that people are different and have different ways of learning and assimilating information. From what I have experienced, the biggest challenges in learning are psychological. It is okay to be confused and it is important to find ways of learning that work better for you.
Two exercises that I have found useful in my learning process are:
- Building machine learning algorithms from scratch.
- Using black-box packages to solve some real-world problems.
Doing these exercises together give me both a higher-level understanding of the system, and also the lower-level knowledge of how the system works.