Dr Kedar Kulkarni is an AI researcher who has made numerous algorithmic contributions to mathematical optimization. Currently, he is a senior researcher at IBM Research, where he creates data science solutions for IBM businesses. In addition, he is in charge of a global team of research scientists and software engineers developing data science/AI/optimization solutions for the sustainability of supply chains.

INDIAai interviewed Kedar Kulkarni to get his perspective on artificial intelligence.

It's great to know that you have a diverse educational background. So how did a chemical engineering student discover a career in data science and artificial intelligence?

I want to start with the clarification that, contrary to popular perception, chemical engineering has more to do with 'math, physics and engineering' than 'chemistry'. So, as a chemical engineering undergrad at IIT-Bombay, I did several courses in pure math (calculus, linear algebra) and traditional Chem. Engg. Courses like fluid dynamics, process control & optimization etc., exposed me to different math areas, e.g., linear/nonlinear programming, numerical methods, ordinary/partial differential equations etc. This exposure developed a broad math foundation, which I believe is very important for a data science / AI career.

After my undergrad, I earned my PhD on a topic related to algorithmic and applied 'mathematical optimization' - which is at the heart of any machine learning / AI approach. This period is where I got a chance to develop and sharpen my skills in algorithms and programming, which is also very important for a career in data science. Post-PhD, over the last 15 years, I've had the chance to work on different challenging problems in industry research, some of which straddle the more conventional math modelling, optimization, operations research areas and the newer data-science area.

A strong background in math and programming helped me discover my path in the Data Science / AI area. And thanks to the exciting and challenging work that came my way, I got a chance to go deeper into the areas of machine learning and AI. 

There is a difference between pursuing an AI job and doing AI research. Who or what encouraged you to seek a path in artificial intelligence research?

The rigorous training I received during my PhD, followed by the challenging research projects that I did as part of industry research over the last 15 years, has naturally impacted the work I like. In a typical AI job, a data scientist will likely choose one of the umpteen off-the-shelf models/platforms and tweak those to solve the industry problem of interest, with little or no scope for taking a more fundamental approach. However, in AI research, 

• one would take a step back & question the status quo, 
• assess the adequacy of the employed methods/tools in the context of the general problem class,
• pose challenging questions about what is lacking in the current approaches, and 
• find what's the best way to address these gaps. 

The above way of thinking has always been attractive to me, and thanks to my training and experience, I am thankfully able to do more of it!

As a graduate in chemical engineering, can you describe the early difficulties you encountered in AI research?

As I mentioned earlier, the different math courses and research projects I did during my undergrad and PhD helped me develop a strong foundation in math and programming. After that, however, I realized I had to skill up in traditional computer science areas like data structures, algorithms & complexity, advanced statistics etc., for me to be more effective. Other than that, 

• I also had to skill up on modern source code management/versioning (read GitHub) and 
• standard software engineering practices (e.g., test-driven development, writing human-readable code etc.) 

since some of these modern tools were not even around or popular when I finished my PhD back in 2007! 

Could you briefly describe the nature of your jobs as a research scientist and data scientist?

​​As a data scientist, my end goal was typically to design and develop a solution that will be deployed into production (at least notionally) soon. And the burning motivation was: how soon can we ship a good enough answer that solves the problem effectively and satisfactorily addresses client needs, and we can generate newer / better versions over time, as necessary. On the other hand, as a research scientist, one typically takes a more fundamental approach, where we try to understand the following: are there gaps in the state-of-the-art to solve a problem of interest, and how we can push the state-of-the-art by devising a new approach or an algorithm that is more generalizable, efficient, scalable etc. all in the industry context.

As a data science mentor, could you share some attractive data science solutions you've encountered with us?

I was a data-science mentor back in 2015, when the 'data science' field had just begun capturing industry imagination, and everyone was curious. I worked with six international mentees to help them get started in data science. Some were students finishing a PhD in a different area, interested in the data science world, while others were working professionals wanting to switch to data science. We followed a course curriculum where mentees would report on their weekly reading/programming assignments and discuss any difficulties they encountered with me. Since they were beginners, I focused on helping them with their fundamentals (math/statistics) and tips on good coding practices. Towards the end of their course, they also did a capstone project where they chose one specific problem of interest and developed an end-to-end ML solution. Most of the datasets they used for these projects were from sources like 

• the UCI repository (e.g., Boston House Prices dataset, Pima Indian Diabetes dataset) and
• Kaggle (e.g., fake news prediction) 

to develop solutions, e.g., regression, classification, and clustering problems in machine learning. 

What is the one skill or quality that a person must have to pursue a career in AI?

It is tough to call out one single skill. However, in my opinion, the following 'hard' skills are essential: math, statistics, programming and algorithms. At the same time, the following 'soft' skills are crucial too:

• Curiosity.
• Critical thinking.
• The ability to stay the course.
• The ability to be mindful of the bigger picture. 

It's great to know that you've developed novel cognitive/analytics/data science/optimization/math modelling solutions for several IBM businesses. Could you perhaps describe briefly some of the problems you encountered and the solutions you provided?

While I have worked on several exciting and challenging problems so far, the following is a recent example where I got a chance to fuse math optimization and ML concepts in a single solution.

In retail, stores typically 'markdown' (discount) the price of products that are not selling well, especially those with a sell-by date (e.g. a new fashion design, a perishable commodity etc.). And surprisingly, we found out that most retail companies have a manual system based on simple business rules configured by an SME to decide on the markdown of several thousands of retail products! However, this system had the challenge of scale – it was becoming extremely challenging to tune and maintain rules for such a wide variety of products. Moreover, the performance of this manual system was not reliable, leading to frequent revenue losses. Thus, it was necessary to build an automated, data-driven, algorithmic price markdown optimization engine that would consistently recommend the optimal markdown strategy irrespective of product type, store location/geography etc. I led the team to conceptualize, design and develop a data-driven / ML-based markdown optimization solution that consistently generated superior revenue estimates than the manual baseline. We also presented this work at the INFORMS 2020 annual meeting.

Could you recommend any academic papers or books that every student of artificial intelligence/machine learning should read?

There are several good resources, so it is tough to list all of them! On the one hand, there are classic textbooks like 

• 'The Elements of Statistical Learning by Trevor Hastie, Robert Tibshirani, Jerome Friedman and
• Pattern Recognition and Machine Learning by Christopher Bishop. 

At the same time, there are several good courses for AI, machine learning and deep learning on various platforms like Datacamp, Coursera, edX etc. 

Is programming knowledge required for artificial intelligence? What advice would you give an aspiring data scientist?

Yes, programming knowledge is critical for an aspiring data scientist to be successful. While in-depth knowledge of AI theory is essential, programming is necessary for the data scientist to translate this knowledge into practical applications. My advice to an aspiring data scientist would be the following:

• Enrol for a beginner course, e.g., on Coursera. For example, Andrew Ng's introductory course on machine learning is excellent.
• Learn Python, the de facto programming language of data science.
• 'Learn by doing': Choose one of the freely available datasets for ML, assign yourself a machine learning problem and independently develop an end to end solution through programming.
• Start participating in data science hackathons and community discussions. Volunteer to present your work and get feedback at a local meet-up.
• If you are feeling adventurous, sign up for a competition on Kaggle, which also has an excellent set of resources for data scientists. One can learn a lot even by studying and implementing the code submitted by past competition winners.