I liked how it was presented in different ways between the book and the video. In the video, they worked on tabular data actually in Excel, and in the book, they did MNIST digit matching from the ground up.

The course: Practical Deep Learning for Coders – 3: Neural net foundations (fast.ai)

My self-imposed “homework” took a dataset of bank fraud and created a Neural Net based on that data: fastai_course/ch4_from_the_ground_up_bank_fraud.ipynb at main · JonathonCwik/fastai_course (github.com)

It was interesting making it from the ground up and with a very simple NN I was getting ~99% accuracy which is great. The actual output was easy enough, it’s either fraud or not fraud (so zero or 1). I’m interested in applying this to timeseries data, and a couple sports prediction things but I’m going to hold off for later.