Coursera vs Udacity for Machine Learning

2018 is an exciting time for students of machine learning. There is a wealth of readily available educational materials, and the industry’s importance only continues to grow. That said, with so many easily accessible resources, choosing the right fit for your interests can be difficult. To help those considering entering the machine learning world, I’d like to share my experience from two courses I took in 2017: Coursera’s Machine Learning course and Udacity’s Machine Learning Engineer Nanodegree program. I found both courses to be very instructive and worthwhile, but very different in nature. If you don’t have time to take both then hopefully this post can help you decide which one is best for you.

Coursera

Coursera’s Machine Learning course is the “OG” machine learning course. Led by famed Stanford Professor Andrew Ng, this course feels like a college course with a syllabus, weekly schedule, and standard lectures. The college feel extends to the curriculum as well. Here is an example slide:

Courtesy of Coursera

If that scared you, you aren’t alone. I usually shy away from courses heavy in math, but I actually appreciated the approach in this course. The course begins with a linear algebra refresher and explains machine learning concepts like gradient descent, cost function, regularization, etc. along the way. It is structured better than any in person college course I ever attended. The material isn’t easy, but that’s a good thing. You come away from the course with the satisfaction of genuinely understanding machine learning, enough so that you could even build your own machine learning framework from scratch.

Udacity

Courtesy of Udacity

Udacity’s Machine Learning Engineer Nanodegree program is the trade school alternative to Coursera’s academia. From basic statistics to full-fledged deep learning, Udacity teaches you a plethora of industry standard techniques to complete the program’s well-crafted projects. The projects are so good, in fact, that I forked their repos on Github and left my solutions up as portfolio items. The final step of the program is to complete a capstone project of your own choosing. While you could theoretically do a similar project on your own, I found the desire to complete my Nanodegree to be a strong motivator; I ended up putting in much more time and effort than I normally would have put into an independent side project. Ultimately, I ended up creating something of which I am truly proud. Udacity’s program doesn’t so much teach as it does provide a framework and motivation for you to teach yourself.

Comparison

Now that I’ve introduced the two programs, I’ll highlight the strengths and weakness of each across a number of categories.

Programming Environment

As I mentioned, Coursera is the “OG” machine learning course; so, it should come as no surprise that the it’s taught in the “OG” 3D math language and programming environment: Matlab. Due to Matlab’s cost and licensing issues, the machine learning world has mostly moved to Python. This move severely limits the utility of the programming assignments because you’ll have to relearn a lot of that work in Python. If you are a seasoned programmer who knows many languages, that might not be a big deal. However, if you are relatively new to programming then this detour may cost you a lot of time.

The Udacity course is taught in a modern Python environment with popular frameworks like Sklearn, Tensorflow, and Keras. The course even teaches students how to use AWS to deploy machine learning software to the cloud. The course also simplifies the process of installing machine learning dependencies with a Docker image and AMI (Amazon Machine Image) for local and AWS development respectively. In fact, the entire Udacity environment is in line with industry best practices and students who learn it will be well equipped in the job market.

Winner = Udacity

Lectures

Coursera’s Machine Learning course was created and taught by the AI godfather himself: Andrew Ng. And this course has contributed in no small part to his reputation within the industry. The lectures follow a single uniform format and each one builds upon the last in a methodical way. Not to mention, he leads every one himself. Lastly, Professor Ng is also very encouraging in his videos, which I thought was a nice touch.

Udacity’s lectures, by contrast, featured a rotating cast of characters, which can create very jarring transitions between sections. I counted at least seven different people lecturing throughout the program. While Udacity attempts to provide multiple content sources for its students, the lack of homogeneity definitely dented my enthusiasm for the lectures. By the end of the program I just skipped right to the projects and watched the lectures, or even searched Youtube, as needed.

Winner = Coursera

Projects

Coursera’s course has programming assignments in which student’s submit code to be tested against automated unit tests. While this model helps the class scale, it leaves you hunting through the forums when things go wrong. That said, I never hit any major roadblocks. The assignments themselves were directly related to the course material and reinforced the lectures. Sometimes it felt like I was actually creating my own machine learning framework; at other times, however, it felt like I was just implementing methods until the unit tests passed.

Udacity’s projects were extremely well designed. In fact, they constituted some of the best educational materials I’ve ever encountered. Each project covered a subject, such as unsupervised learning, reinforcement learning, linear regression, in which you solve a multi-step machine learning problem and write about your approach and understanding. When you feel that you have completed a project, you submit it to be graded by a HUMAN. The quality of the feedback that I got was incredible. The final project is a capstone that you get to pick yourself, but it is still reviewed by Udacity’s staff. The proposal and final report ended up being one of the best portfolio items I have ever created and one of the things I am most proud of in my programming career.

Winner = Udacity

Cost

Coursera’s price is hard to beat because it’s free. To get the certification its $80. If you are machine learning on a budget then Coursera is a great choice.

Udacity has recently changed its pricing model for the Machine Learning Nanodegree. When I entered the program, it was $200 a month. Now it is a $999 flat fee. The per month pricing model incentivized me to finish the program quickly in only three months. Though I must admit, given the quality of instructor feedback, even with the price hike tuition still seems reasonable. The highly-skilled labor that is meticulously reviewing projects can’t pay for itself. With such a high dollar amount, however, signing up for the Nanodegree program is obviously a much bigger consideration.

Winner = Coursera

Conclusion

While the courses tied on the number categories won, I am going to pick a winner. It is…

Udacity. It may come as no surprise that a paid course beats out a free one, but the Udacity Machine Learning Engineer Nanodegree program gave me the confidence to professional pursue machine learning positions and opportunities; and for that, its entry fee was a very small price to pay. That said, I would still recommend you do both courses. Start with Coursera, so that when you use “batteries included” high level frameworks, you understand the low level details and have a better appreciation of what you’re actually coding. After you’ve built a strong conceptual foundation, further refine your skills by learning practical, industry standard practices at Udacity. Overall, I am so glad I took concrete steps to enter the machine learning world in 2017, and I would encourage you to do the same in 2018.

Coursera’s Machine Learning Certificate

Machine Learning Engineer Nanodegree Certificate