Recently I completed Statistics and Data Science MicroMasters successfully. The program is well designed to deepen knowledge and skills required in data science. I’d like to share how good it is and how hard it was.
· Who I am
· Why I took this program
∘ Probability — The Science of Uncertainty and Data (6.431x)
∘ Machine Learning with Python: from Linear Models to Deep Learning (6.86x)
∘ Fundamentals of Statistics (18.6501x)
∘ Capstone Exam in Statistics and Data Science (DS.CFx)
∘ Data Analysis: Statistical Modeling and Computation in Applications
· End thoughts
My name is Teru Watanabe. I’ve been working as a data scientist for five years and before that, was a software engineer for ten years. Although I’ve always been in a technical role, I don’t have any formal education related to these fields. I’m a self learner. When I get curious about something new, I just learn it on my own. At first I just learn it for fun, then after a couple of years I take on a new role related to the new field. If you are interested, you can find more details in my Linkedin page.
I found this program at the beginning of 2019 while searching for online resources to advance my knowledge and skills in data science. Since it was from the prestigious school MIT, I jumped in to take one course which happened to be just starting that time. The course was Fundamentals of Statistics (18.6501x), which would turn out to be the most challenging not only in this program but also of all the courses I’ve ever taken. I tried the first few weeks but it was very hard to follow because of the level of mathematical rigorousness that the course has as the basis. So I decided to drop out.
I was shocked that I could not even understand the content from a statistics course described as “fundamental”. At that time I had already been working as a data scientist for three years and completed several statistics courses. “Oh my, I still don’t know something fundamental in statistics.“ A little disappointment and an expectation that I might be able to deepen my understanding of some fundamental aspects of statistics gave me the initial momentum.
After dropping 18.6501x I started to take Probability — The Science of Uncertainty and Data (6.431x) in May 2019, which is the recommended first course in the program. It took me one and a half years to complete all the courses and the final exam. I was not so committed to this long journey in the beginning, but as I proceeded I got more convinced that this program teaches me something I didn’t have and ended up taking all of them.
Statistics and Data Science MicroMasters is a series of graduate level courses that MIT provides through edX platform. The program comprises four online courses and a virtually proctored exam. Each course is semester-long (3 to 4 months) and requires at least 10 hours every week. It’s a MOOC but the pace is basically the same as the on-campus version. Every week they release a lecture and it contains:
- videos equivalent to two classes (~180 mins)
- exercises (counted towards the final grade)
- homework (counted towards the final grade)
Exercises and homework are difficult. They check if learners understand the contents deeply and can utilize the knowledge and techniques. They also have deadlines (usually 2 weeks from the release) and limitations on how many times you can submit the answer (usually from 1 to 3 times per problem). Both rules are strict. If you don’t meet these criteria, you don’t get the credit.
Because of the restricted number of submissions, sometimes I had to contemplate for a long time before submitting my answer. This design forced me to think deeply about the topics I learned. The good news is that when I was really stuck in a difficult problem, I could ask for help in the discussion forum. There were always TAs and classmates who pointed me in the right direction. For exercises and homework, discussing how to approach problems is allowed as long as we don’t reveal the answers directly.
Aside from weekly lectures, each course has midterm and final exams. They are open-book exams but not that easy that you can just search for answers from lecture materials. Most of them are hands-on calculations and require precise answers to get the credit. In contrast to exercises and homework any discussion in the forum is not allowed and no immediate feedback for the submission is given for exams. So I needed a thorough review before taking an exam every time. This is also a good design that gave me a chance to deepen my understanding.
The following is the summary of each course. Difficulty and effort are based on my personal experience.
Probability — The Science of Uncertainty and Data (6.431x)
Effort: 15~30 hours/week
Length: 16 weeks (May 2019 ~ Aug 2019)
A lot of hands-on calculation of different types of probability and statistics. It has a good balance between mathematical formalism and practical intuition. Prof. John Tsitsiklis is really good at explaining complicated concepts in an intuitive way. Watching his lecture videos is fun. The content is sophisticated because it is a long-running course (more than 50 years!) in MIT on campus and has been refined throughout the years. By completing this course, I could develop a mathematical foundation for framing statistical problems more formally. This course equipped me with required knowledge and skills to tackle 18.6501x.
Data Analysis for Social Scientists (14.310x) & Data Analysis in Social Science — Assessing Your Knowledge (14.310Fx)
Effort: 10~25 hours/week
Length: 11 weeks + 1 week (Sep 2019 ~ Dec 2019)
A series of practical data analysis with R programming in social science context. It covers some theoretical backgrounds but the emphasis is more on their applications. It’s mostly about hypothesis testing and causal inference in the real world. All techniques are introduced with some actual applications. It teaches us why and how a specific technique is used by walking through data analysis performed in famous papers. This course was relatively easy for me because 1) the first half (preliminary for the latter) overlaps with the contents from 6.431x and 2) I was already familiar with R programming and could skip DIY-style R programming contents. (R programming is not the main focus in this course.)
The course is separated into the content part (14.310x) and the exam part (14.310Fx). You don’t have to take the content part as long as you can pass the exam part. I wonder what kind of people can entirely skip the content part and pass the exam part though.
Machine Learning with Python: from Linear Models to Deep Learning (6.86x)
Effort: 10~25 hours/week
Length: 15 weeks (Feb 2020 ~ Apr 2020)
Low-level understanding of machine learning with Python programming. It’s Python and machine learning but not how to use the high-level API of scikit-learn. The focus is how machine learning algorithms work under the hood. It’s a step-by-step walkthrough of algorithms with linear algebra and calculus. The course doesn’t cover many different algorithms though, it teaches us the most important underlying techniques used in many algorithms. This course is more on the theoretical side and a good place to advance your understanding of machine learning.
Fundamentals of Statistics (18.6501x)
Effort: 20~40 hours/week
Length: 15 weeks (May 2020 ~ Aug 2020)
The hardest course ever. It’s a mathematical statistics course. It teaches us formal mathematics behind statistics and their applications. It was the most mathematically rigorous course I had ever taken. Prof. Philippe Rigollet is excellent, but some concepts were really hard to digest because of their nature. I encountered many new, difficult concepts and performed endless hands-on calculations by utilizing them. By going through the hardship, I became comfortable with following mathematical explanations in papers and books. The hardest and the most rewarding course.
Capstone Exam in Statistics and Data Science (DS.CFx)
Effort: 2 hours x 4 exams
Length: 2 weeks (Oct 2020)
This is the final exam. It consists of four two-hour exams and we need to score 50% as the total of the four exams to graduate from the program. It’s closed-book and strictly monitored by proctoring software. What’s hard about this exam is that I had to review all the contents covered by all four courses. It took me a month of hard work to complete the review. The exams are less difficult than I imagined from hindsight, but if I didn’t go through the hard work it would have been a disaster.
Data Analysis: Statistical Modeling and Computation in Applications
Length: 15 weeks
While I was taking the program, this course did not exist. It is a newly released elective course that you can select as an alternative to 14.310Fx. The first batch is going to start from Feb 2021. According to the course description page, the focus seems to be more on hands-on experience with Python programming.
Aside from the knowledge and skills that I could acquire, I got the following benefits after my graduation.
- I became an official affiliate of MIT and joined the alumni network.
- I got the privilege to apply for PhD programs in MIT.
- I got the privilege to apply for Master programs provided by some universities all over the world.
The list of the graduate programs credential holders can apply is here.
While I have a full time job, this much of time investment was very hard. Not only nights after work on weekdays, but also countless full commitments on weekends. It was a long, challenging path, but it’s worth it considering what I’ve learned.
I’d like to say thank you to JDSC which gave me time to study. The company promotes employees’ skill advancement. All of us (including non-technical positions) in JDSC can use two days every month to study anything based on our own will and interest. This system helped me a lot when I had to prepare for the final exam in a limited time duration.
By the way, we are hiring.