Based on 6 years of experience in ML
Disclaimer: 2 years of academics + 4 years of professional experience. I don't claim to be a mentor/coach, and I don't claim to myself for having an extraordinary track record. But whatever I put in this blog post is the result of my real-life experience of interviewing over 100 profiles in the ML domain in the last 2-3 years.
What we are witnessing today is the surge in machine learning courses and the enormous “interest” for undergraduate students pursuing careers in ML. Personally, I am approached by many undergraduates and even experienced people who are looking for guidance on how to start with machine learning work. In this blog, I integrate ideas and surface the myths that a general audience has as they begin their journey.
My journey
- Not fresh in ML: When I began my professional journey, I was a fresh undergraduate student, but not fresh in the ML domain. I have been working on projects at NLP for the past two years at university. I realized that I have done some of the online courses, mentioning that people are very familiar with the domain. I strongly oppose this concept of learning. Starting with an online course is a good thing and perhaps essentialI did the same thing too), stop there. There are no viable ideas (read the section below).
- ML jobs were not provided with Silver Spoon: I had a bit of ML experience, but getting a practical ML job was not easy. I have always wanted practical exposure in my full-time role and have not leaned towards an internship/higher research/RA position. Luckily I got it and decided to stop offerings from the international MNC that I received in the university placement. TL;DR: I wanted to work in the ML domain, so I was ready to let go of a brand that everyone would want to work in some capacity on another day.
- Academic ML vs. Production ML: Certainly, many have already learned that academic machine learning is very different from production machine learning. I can't say anything more than say this is the biggest truth (please read the detailed section on this in the next part of the blog).
Invasion of machine learning – What I've heard
Don't be shaken by the hype the Field has, trust the rationale behind why you want to try ML. Here are some pointers you should think about.
- Online Course: This is a great starting point and there are plenty of resources available today. The downside of this is that there are many candidates who have done one or more of these online courses. What distinguishes you from others?
- ML Project: This is an immediate response that many people give. People with mild projects talk about sentiment classification and housing price predictions, while deep learning enthusiasts talk about text generation projects and image classification using ResNet or similar architectures. Again the same question, Many competing candidates have also done the exact same project, but what are the other candidates?
- Deep learning keywords only: This is another group of candidates who have found themselves spontaneous in all the buzz words of the deep learning ecosystem. To quote Buzz, do you honestly think you stand out from others?
Invasion of machine learning – What I suggest
Continuing the pointers in the above sections, here are some guidelines for developing profiles in the machine learning domain:
course
In my opinion, the course needs to understand the rationale and concepts behind ML algorithms.
- Although courses range from shallow to advanced courses, I always prefer to gather information from formal courses/textbooks rather than crash courses. The crash course helps you quickly point out outlines and bullet points. Formal courses can provide deeper insights into some of these pointers and textbooks. At the same time as dots, it helps to surface many of the concepts not explained in online/formal courses. Always prefer quality over quantity! ML is vast and no one expects you to know everything.
project
Again, a gentle project or project (issues) in an online course is a good way to start with actual implementation, and of course not enough if it was only done within the scope of completion. What you do in addition to a particular project will be counted. Here's how to extend your project:
- Have you hosted it on github with the appropriate readme? This solves two objectives, provides visibility into your work, organising it, and helps maintain implementation at your fingertips whenever you want to refer to it.
- Have you conducted an extensive analysis of the experimental results? __ From EDA to cost functions, data, hyperparameters, and error analysis. For example, if you train a deep learning system, will you be able to interpret the results? Can I have plugin code to visualize attention maps?
- Have you applied what you learned from other data? When I learned about emotional classification, I learned how to perform textual classification primarily. Have you applied the text classification you just learned in another dataset: imbalanced dataset, multi-class dataset with many classes, Big-Data, etc. What differences should you pay attention to WRT differences regarding text classification of sentiment data? Can you do/learn anything about the differences you see?
Again, I always prefer quality over quantity! 2–3 great projects stand out from 10-15 generic projects
Deep Learning and Statistical Machine Learning
If you are a beginner in statistical machine learning with lots of information about deep learning, you have a “high” chance to do do not have They take precedence over candidates that are strong in statistical concepts.
- Make that the point Statistical machine learning is extremely important to know and will definitely be asked in ML interviews. Additionally, if you are working in the field of generic ML (tabular, numerical, categorical data) or NLP (natural language processing).
- Deep learning is more than just knowing about CNN, LSTM and transformers. It also includes the emergence of architectures that can work well for data and business issues. This is something to learn and develop. One of the best ways is to read research papers to learn how to build a DL architecture. Therefore, you should not be confident when you are training some models using CNNS/LSTMS/transformers.
Again, I always prefer quality over quantity! A good understanding of some DL architectures and some statistical ML models is far better than just touching on everything.
I know what and why
There are many accepted structures that we've heard from candidates that are mostly true, but it doesn't explain why those components work.
- why Does SVM performance decrease as dataset size increases? Which aspect of the model means “poor”?
- why Is bias used in neural networks? Is it used only in neural networks?
- why Is batch normalization used? How is batch normalization achieved?
- why Is multi-agent games difficult?
Answers to these questions can only be observed/understanded if you dig deeper into concepts and readings. This is where formal reading and true interest can be useful. PS: You are not expected to recognize everything, but a small start is still a start.
The final words
- Start small – ML is vast and don't be overwhelmed by the available resources and domain depth. Choose one and complete one.
- Develop width and dive deeper – Do not limit yourself to deep learning or table data only during the start. At first, explore a little more, understand the space, and gradually dive deeper. The same can be said about algorithms. First select 4-5 algorithms and then start digging them even deeper.
- Specialize – In my opinion it's always better to dig deep into some domains. There are multiple indexes to select. Monitored and unsurveillanced, NLP vs CV vs. image, structured vs. unstructured. After the initial width, it is very important to begin developing strong profiles in several specialties based on your interests. It's not harmful to notice others, But it definitely has superficial knowledge of everything, and there's not one enough for one.
- Read research papers and blogs – Reading related research papers and blogs (medium blogs, company blogs, Kaggle kernels) can help you understand the field of problem, develop different perspectives, recognize different ML problems, and learn much faster than traditional methods of applying ML in idealistic scenarios.
- I'm never confident – You don't believe it, but ML is so vast that you can't claim to be a star until you're alone! It's humble to always learn more from everyone. There is always something else that others know you won't do.
About me:
