Loading
Loss function first. Model second.
the shape of this subject
ML has a copy-paste culture that hides misunderstanding. Brainback keeps you defining the loss and the evaluation metric before you tune a hyperparameter.
Because machine learning rewards the students who do the reps — and punishes the ones who outsource them.
Add Machine learning as a Learn subject and every problem in these areas routes to the Socratic tutor — attempt-first, hint-laddered, output-guarded.
Attempt → correction → guiding question. No answer. Ever. Verbatim from a beta session.
Logistic regression with L2 penalty; step size 0.01…
What loss are you minimizing? What does its gradient tell you?
no answer · one guiding question · your next move
the anti-pattern
Answer machines make machine learning look like a solved problem. It isn't. The exam room reveals whose reasoning was built and whose was rented.
the recall card
Every wrong attempt and low-scored explain-back becomes a card. SM-2 schedules them — the tricky ones come back tomorrow, the solid ones drift out to next week, then next month.
Recall the core concept in your own words. Then apply it to a fresh example under 2 minutes.
The subject where the shortcut has always been most expensive.
Limits, rates, and shapes of change — earned, not delivered.
Where symbolic manipulation becomes second nature — or doesn't.
See the diagram. Not just look at it.
Unit circle, but also unit intuition.
Vectors are directions, not numbers.
→machine learning — ready when you are
Attempt-first. Never the answer. Add it as a Learn subject in 30 seconds.