Past ML Courses
I’ve tried a few of the popular machine learning courses over the years. I started with Andrew Ng’s Machine Learning on Coursera. The lectures were great, and gave a solid introduction to some of the foundations and math behind ML. It was in-depth, but the pacing was a bit slow - weeks spent on basic models like linear and logistic regression. Also, the programming environment was MATLAB-based, both unfamiliar to me, and dated as an ML stack. With that as an excuse, I didn’t follow through with the programming assignments. While I wasn’t personally engaged, the course was still incredibly popular, and a success story that helped drive the rise of Massive Open Online Courses (MOOCs).
I was excited to see a new course that focused on a more pragmatic approach to Machine Learning. I started the fast.ai course last week, and have been incredibly impressed with both the course material, and the general pedagogical approach. For the first assignment, students implement a state-of-the-art image classification algorithm, and submit real results to the data science competition site Kaggle. A great deal of work was put into abstracting out complexity - the entire neural network implementation is only 7 lines of Python code. I understood only a little of what was happening under the hood (and that was just from past experience and background knowledge - must be overwhelming coming in blind). But I was happy to see that in subsequent lessons, the “magic” was explained as we delved into the underlying Keras framework.
Reflections from a Former MOOC Creator
From my experience helping launch and run BUx’s Sabermetrics 101 MOOC with Professor Andy Andres, I know how powerful it is to give students an immediate sense of accomplishment, especially for intimidating subjects. While MOOCs are often praised for their openness and accessiblity, they have much lower completion rates than tquick engagement approach was keraditional, in-person courses. 10% is the golden standard. In the first week of the Sabermetrics MOOC, we had students (many with no programming running SQL queries to calculate baseball statistics. We surpassed the 10% completion rate by a fair amount, and this method of quick engagement was key. Knowing how difficult it was to set up an accessible SQL environemnt, I applaud Jeremy for doing this for the far more complex task of running a neural network.