Interim update, written during the term.
Deep learning without the exam: why I rebuilt the course book as a website
This term I'm sitting in on the Deep Learning lecture at HPI. To really understand the material, I rewrote the first part of the course book as a simplified German study site with interactive demos.
You understand the formula for gradient descent eventually. What a learning rate that’s too large does, you only understand once you watch the point jump out of the valley.
This term I’m sitting in on Prof. Christoph Lippert’s Deep Learning lecture at HPI, held in English. I’m only listening. I’m not taking the exam and I don’t get credits for it. I still want to know the material. So I started rewriting the first part of the course book, in German, simplified, as a small website with demos you can play with.
So far it’s eleven pages. Three lessons: linear regression with gradient descent, logistic regression, and then optimisation and regularisation. Plus labs on optimisation and regularisation and a few extras: exact solutions, curvature and convergence. KaTeX sets the formulas, and there’s a glossary.
Here’s the basic idea the demos make visible, in words. You have a loss curve, a valley, and a point somewhere on the slope. Gradient descent checks which direction goes downhill and takes a step. The learning rate decides how big that step is. If it’s small, the point crawls into the valley. If it’s right, it gets down quickly. If it’s too big, the point jumps across the valley, lands higher on the other side than before, and swings itself up. Regularisation is just as visual: the stronger it is, the more large weights get punished, and the simpler the model gets.
Why the effort? Because explaining something shows me where I haven’t understood it. Building a demo forces me to translate every line of the formula into code.
The whole thing is plain HTML and JavaScript, no framework, no server. It isn’t published; so far it’s a study aid for me. I know the pattern from myself: with ¡Vamos! for Spanish it was the same, build a tool first, then learn with it.
What I’m taking away so far: you learn differently without an exam. No pressure from a grade, and instead the freedom to stay on one single point until it sticks. And interactive demos are the fastest way for me to get a feel for a formula.
Tools and skills: HTML, JavaScript, KaTeX, gradient descent, loss functions, logistic regression, regularisation and optimisation, plus didactics, meaning the question of how to explain something so that someone without background gets it.