Week notes — 12th September 2024

I’ve started going through Brunton & Kutz’s Data Driven Science and Engineering book, taught as a series of YouTube lectures. I still have to do a final project for the NLP course I’ve been doing. I’m waiting for the right moment to combine that final project with another project I might be doing.

In the meantime, I’m looking at some Engineering ML. The course focuses on:

“…bring[ing] together machine learning, engineering mathematics, and mathematical physics to integrate modelling and control of dynamical systems with modern methods in data science.”

The YouTube lectures are incredibly well done, and ‘homework’ PDFs with code (Python & MatLab) and datasets are also provided.

Why engineering ML? I’ve been loosely modelling my ML learning on what you might be taught on a master’s program, where students are typically exposed to a range of topics: NLP, computer vision, ML for graphs, stats for ML, etc. Mirroring that, I’m trying to spread myself out a bit.

Engineering ML specifically appealed to me because I’ve always been fascinated by engineering and manufacturing — and in some places it connects to my dusty undergrad physics skills.

NLP courses, perhaps because of the underlying messiness of language, tend to take a ‘cookbook’ style approach, giving you a set of methods that work, with less explanation of why.

At least so far, Brunton and Kutz’s approach feels very different. I haven’t looked at Singular Value Decomposition before, and it’s already giving me a different appreciation of PCA. Understanding one topic from two different directions gives a much deeper sense of understanding.

Brunton & Kutz strongly emphasise spatial intuition, specifically the idea of solving problems by changing the coordinate system, which I find appealing and intuitive.

I intend to forge ahead with Brunton & Kutz for the next two weeks and see how it feels. I will return to do some project work, probably deploying BERT, when I have a suitable problem space to explore.