Raahul Seshadri

It’s helpful to talk about the impact of your decisions or other factors rather than the “cause” of it. For example, instead of talking about an “imbalanced dataset,” talk about how your model will in-accurately predict that “every 10 people out of 100 will be erroneously rejected on their loan application.”

Things have been given many different names today, just like Data Science. But one must still focus on getting the basics of statistics in place. The start might be a bit slow, but it snowballs into an enormous repository of knowledge soon. If you have your basics right, you don’t have to “learn from scratch” when something new comes along. You can just give it a cursory read and “you’ll know it.”

Whenever I want to look for courses, I just type the name of the topic (say, matrix factorisation) and then add a “site:.edu” to the Google search. This gives me a list of university websites that cover these topics. Many of them have video courses, or at least notes, and plenty of references to academic papers and authoritative textbooks. That allows me to declutter my first page of Google and get to learning the first principles without fuss.

Asynchronous courses allow you to learn at your own pace. I prefer written materials (books, notes) that concisely cover what’s to be learnt, and then providing plenty of exercises to practice. One has to get to practicing soon instead of just spending a lot of time reading/watching videos. Or else, instead of practicing data science, we just end by “practicing to read.”

NLTK is a versatile tool to learn to basics of natural language processing. In the world of LLMs, what you usually see is a black box without a lot of insights into how things work underneath. That’s where NLTK, because of its excellent documentation, can help be a good primer into the world of NLP. After the basics are mastered, then you can expand your scope to other tools and more sophisticated methods.

When working with any NLP project, consider what the rest of the tech stack is going to be. Python is a very popular choice as a language not just for data science, but also for building web-apps. Thus, it makes a lot of sense to use Python as your frontend language. ML frameworks usually use a different backend language, like C++ or Rust, to make sure you don’t compromise on performance.