Sagar Setu is a co-founder and chief technology officer (CTO) at Milky Way AI. He graduated from the Indian Institute of Technology Kanpur with an M.Tech Dual Degree in Aerospace Engineering and a PhD in Aerospace Engineering. 

NCERT, India, has awarded Sagar the National Talent Search Examination Scholarship. In addition, he is the recipient of a Boeing-IITK scholarship.

Sagar has improved a physics-based simulation model using neural networks for the autonomous landing of a helicopter after an engine failure.

INDIAai interviewed Sagar to get his perspective on AI.

How did you get started using AI as an Aerospace Engineering researcher? How did this transition happen? 

My doctoral research was on developing a high-fidelity rotorcraft simulation. It is the testing control system for an unpowered rotor (failed engine) scenario. However, low mathematical complexity physics-based models cannot predict how the rotor behaves in descending flight, and more complicated models are challenging to execute in real-time. There was a need to fill the gap in the physics-based understanding with data. We observed that traditional system-identification techniques required a complicated set of maneuvers to fill this gap. However, by carefully crafting a neural network which can augment the physics-based model, We significantly reduced the complexity of the experiments. That's when I realized the AI field's power for research in other disciplines. 

In your role as a robotics researcher in Singapore, what were some of the most exciting problems you faced? 

Unlike outdoor flight, working with indoor drones poses a challenge: the absence of a reliable GPS signal. Using visual cues to navigate (like a human being) is an active field of research, and engineering such a solution requires a balance between flexibility, reliability and computational power requirements. Nevertheless, indoor navigation is a field that I still find exciting and has remarkably more possibilities with the current state of AI and computing technology. 

Tell us about some unique difficulties you encountered when developing IoT technologies for managing natural and civil water resources as a co-founder. 

We worked with academic institutes in our early days to collect real-time water levels from natural water systems. One of the unique challenges was to ensure the security of the instruments deployed in remote areas. To solve this, our team worked to engage the local community, helping them understand what the instruments do and how the data collected will be used by researchers to understand the water cycle in the region and eventually solve some of their woes related to flood and drought. This shows the importance of involving the wider community and getting them acquainted with how any research is beneficial to them, considering them as an essential stakeholder whose buy-in is always required in a project. 

How do you use AI as a co-founder of Milky Way AI? 

Our core technology is product recognition, i.e., identifying a product in an image to its most granular level of distinction. It differs from a traditional object detection pipeline two-fold:

a) It requires both coarse and fine-grained recognition. For example, we need to identify all kinds of retail products, whether a bottle of soft drinks or a bag of chips. At the same time, we identify the exact flavor of the bag of chips.

b) The dataset is constantly drifting. New products are launched daily, and existing products also have their variations temporarily out in the market. For example, there might be a Diwali special package of a box of chocolates, which is not a new product, but a promotional variation of the same product.  

Both these challenges pose exciting constraints on how we can use the understanding of traditional object detection and image classification and what components need to be explicitly re-designed for this kind of problem. 

How do you ensure that your platform follows ethical rules regarding AI and applications of AI? 

Although we do not deal with sensitive data, we constantly strive to modularize our AI into more explainable pieces. In addition, we follow the best cybersecurity practices and take constant feedback from a team of human verifiers who can point out any bias in the system. 

What will be the next big thing in AI research during the next decade? 

A big hurdle in the broader adoption of AI is the cost associated with it, which comes in two forms –

a) Requirement of a large amount of data for training the models and

b) Computationally cost of training, re-training and running AI models.

We are one of the limited companies worldwide that have successfully adopted few-shot learning, requiring fewer data and minimal model re-training to make the process economical. However, much more needs to be done to reduce the cost associated with initial training and ongoing running costs. I am hopeful that these areas will majorly improve in the next decade. 

In your perspective, what are the primary areas where Indian researchers fall short? 

There are some excellent minds at work in Indian academia in the field of computer science in general and AI in particular. However, I feel there is a lack of an ecosystem to promote quicker iterations and experiments. To instantiate, China is known for its advantage of access to massive amounts of data. Large corporations have access to practically unlimited computing resources to run massive experiments in parallel. 

What advice would you give someone wanting to work in AI research? How can one get ahead the most? 

My primary advice would be to spend some time understanding the domain of the problem. The mass adoption of AI is a recent phenomenon. There could be an inspiration in how humans have been solving problems before that could come in handy while designing problem-specific AI models. Just look at the evolution of AlphaGo, and how it respected the boundaries and appreciated the almost mystically creative minds of the champions of a real-world game that humans have been playing for centuries. Alpha Zero, the version that doesn't use human data, would have been challenging to build without its predecessors and sticking to a problem to solve. 

Could you give me a list of books and articles essential to AI research? 

For readers looking for an introduction to the field, "The Creativity Code: How Ai is Learning to Write, Paint and Think" by Marcus du Sautoy is an excellent read for understanding the difference between traditional decision-making by computers and that by AI. My first recommendation for a more technical introduction, online courses by the revered Dr Andrew Ng. Finally, there is no better read than recent research papers on the topic of severe researchers looking to make a career in the field. However, that requires working on a problem for some time, identifying the specific problem that interests you and finding relevant papers.