Finding your niche in Machine Learning

Why and how to find ‘your’ area

Photo by Tim Foster on Unsplash

Machine learning can be overwhelming.

I was talking to a Research Scientist (a.k.a with a Ph.D.) in my company and she mentioned how she feels insecure about her ML skills. There is so much to do. So much to learn every day. Feeling this way is independent of which career stage you are at. Whether you are starting, in your grad school, or already working in the industry. In any phase of your career, you need to have your niche or ‘your field’.

First, let’s try to understand why having your own niche is important and helps to navigate ML career options.

Online courses don’t add value

Online courses add value to your personal knowledge but not to your resume.

Since today everyone is building the image detection classifier in the Coursera course.

Having it on your CV won’t add much value. On the learning side, these courses are structured and good for starting up. But, eventually, you need to build something of your own to get a better understanding of the topic and getting an edge over other people in interviews.

Niches help you to filter grad school and jobs

I get this question a lot-

How do I decide which grad schools to apply to?

Filter by topics is always my answer. Most Universities offer a wide range of courses on a particular topic. Make sure you filter according to your interests. Look at the courses offered and check if they are something you are interested in (because you are going to invest a lot of time in studying those).

This also helps in filtering the type of jobs you would be a ‘good fit’ for and better your chances of getting in.

Helps you answer what do you work on?

The first haunting question in most interviews I was asked – What do you work on?

How does anyone answer this if they have worked on 10 different topics and know bits and pieces about each one of them?

Most interviewers like to dive deep into one of your projects and talk about it for the rest of the interview.

As an example when I was interviewing for various roles, my experience in Multimodal Learning and Graph Neural Nets helped me in getting my current role even though my work is not directly related to these fields.

Niches eventually become your skill

Everyone is a Data Scientist today. What makes you different is your specialization. If a person knows Active Learning apart from basic topics in ML, he could benefit the team in that particular domain. It separates you from others. All the queries regarding that particular domain get directed towards you and it helps in trust-building within your group.

Now that you are somewhat convinced about why finding your area of expertise is important, let’s look at how anyone can find ‘theirs’. Dr. Hima Lakkaraju, who is an Assistant Professor at Harvard said in a recent interview of hers

Asking for the best topic in ML is like asking if there a way that I can try to optimize my life(or my research topic) such that I win the Turing award? The simple answer is No.

In such a fast-moving field there’s no way to optimize your topics to the most exciting one which will also be useful after years. Trends change over years and it’s a continuous learning process. But here’s what I have followed over the course of years.
Getting topic list- You can find the current ‘hot’ topics from the workshops, tutorials and conferences. As an example, I attended ICML this year and the first thing I did was to look at all the workshops and tutorials taking place. Workshops usually present unpublished work and are great for getting an idea of future trends.

• Learning from tutorials

  One of the best ways to get into a topic I believe is learning from tutorials. They provide a great timeline of the ongoing research and previous work done.
  
  

• Finding code implementations

  I find the widely used libraries in that field and try on some basic code. I then try to skim through the research implementation of some basic papers. This is very personal but until I have done this I don’t feel fully satisfied in getting to know the actual workflow of the topic. Obviously, this differs if you are working on the theoretical side of things.
  
  
• The first project
  
  Find the most basic problem of the field and implement that. This helps to know the problem setup better and get your hands dirty with the code. Let’s say you want to dive into Graph Neural Networks. Try to access the widely used graph data and build the simplest GNN in Numpy or Torch Geometric.
  
  

• Start Networking

  The best way to learn is from the experts themselves. Contact people from the field and ask them relevant questions. That being said don’t send generic emails and spam people, that has never helped anyone. And no one replies to spam. You can ask some questions like ~ ‘Could you comment on the scope of the field’. ‘What are some applications of the topic in the industry’, ‘What is the smallest extension of your proposed approach?’, ‘What are your favorite papers on the topic?’
  
  
• Post-starting up
  Now with everything you have learned, you are good to go with ideas and implementation of your own.