This is how I prefer it. Give me the lay of the land, if you will, then I can dive into what is interesting.

There's nothing wrong with a kitchen sink course. Mostly likely the in-depth deep learning classes have their own course.

I'm taking this now, this year they split it into two classes, and so now there's much much more time to go into theory. But this is still an applications-focus class. They teach you general stuff: What to model, how to build, evaluate and improve models, how to do this on real world data, common pitfalls, how regression and classification works, then some specific models, tradeoffs and similarities between those models, then some methods to improve results from those models.

But this is great, because if you want to learn more in depth about those things, you can just take the second class in the series or a more advanced classes, and / or pick up a book, and you're already at a great position to improve.

I found the lecture useful. I haven't made it to the end, but there is a good visualization on solving the XOR problem without the Kernel Trick, starting at 26:57. She shows a multi-layer perceptron and uses two hyperplanes (hidden layer) to write a remapping so that a single hyperplane (output layer) can classify the 4 points. It is a general perceptron example, so it's only one hidden layer, but it shows the gist of DL (transforming data into a new representation) in a way that I hadn't seen before.

The only prereqs for this course are an introduction to programming course and an introduction to statistics course, so this isn't surprising.

Does the "1" in "109" imply that this is a intro course?