Back to school

I’ve signed up for a data science certificate through Harvard. It’s a part-time program, the kind you do in the evenings with a day job, and it’s going to take a while. I’m a little nervous about it, which I wasn’t expecting.

The honest reason is that I’ve been faking it for a few years. Every product I’ve worked on recently has had some piece of it that was really a statistics problem: a ranking, a forecast, an anomaly alert, a “customers like you also…” feature. Most of the time I’ve been the person in the room who knows enough to say “that’s a regression problem” or “you’re overfitting”, and not quite enough to sit down and prove it. I can read the papers, mostly. When I wrote up the Transformer paper last month I had to stop and work through the softmax scaling on paper, and it took me longer than it should have for someone with a math degree.

That degree was a long time ago. I liked it at the time, and then I spent the following couple of decades writing C++, C# and JavaScript, and the math slowly turned into something I knew I used to know. Linear algebra I’ve kept up with, more or less, because graphics and the odd algorithm keep dragging me back to it. Probability and statistics I’ve mostly lost, except for the parts that come up when you argue about A/B tests.

The thing that actually pushed me was Andrew Ng’s machine learning class back in 2011. I did the whole thing, every exercise, in Octave late at night, and it was one of the most fun things I’ve done as an adult. But it was built to get you productive quickly, and on purpose it skips a lot of the why. I came out of it able to train a model and not really able to tell you when I shouldn’t trust one. Six years of reading blog posts since then hasn’t fixed that. I’d like to understand the stuff properly, with problem sets and deadlines and someone grading me, because I know myself well enough to know I won’t do it otherwise.

Most of the coursework is in R, which is going to be an adjustment. I’ve written exactly one R script in my life, to make a chart, and I remember it being strange in ways I couldn’t quite put my finger on. Everyone who uses it seems to either love it or put up with it. I’ll report back on which camp I end up in.

The other reason, which I didn’t expect to matter as much as it does, is that I’m going to be a beginner again. I spend most of my week as the person people come to with questions. It’ll be good for me to be the one asking them, and probably good for the people I manage to see me struggling with a problem set.

I’ll write some of it up here as I go. Mostly the bits where I get confused, since those are the posts I find most useful from other people.