Learn machine learning · part 6 · session 5
Debug four trainings
Four runs, four faults, one fault each, and no answer key on the page. Everything you need is in the diagnostics, and two of the four share a curve.
After this session you will be able to diagnose a failed training run from its diagnostics and say which number rules out the alternative.
The gate is four of four: every fault named, and for each one the diagnostic that rules out its nearest alternative.
01 · The question
Four runs, four faults, one fault each, and no answer key on the page. I trained all four once and shipped them with their diagnostics, so nothing here waits on your device and you can scrub any of them back and forth.
Two of the eight faults in the list are setup mistakes you have not met yet. Never scaled means one column arrives in thousands and another in fractions, so the large one swamps the . Never shuffled means every row of one class arrives before any row of the other, so the model chases one label and then the other.
Everything you need is in the diagnostics: two loss curves, a weight histogram, the split summary, the permutation importances, the configuration and the gradient norm. Two faults can produce the same flat curve. Pick the number that could only have come from one of them; that is what “the diagnostic that settles it” means. This is the last session of the part and it is meant to be its test. Commit first: how many of the four do you expect to get right?
02 · Predict
Before anything runs
How many of the four will you diagnose correctly?
If you want a line on why, name the one you think will be hardest.
03 · Try it
A column written after the answer
illustrativethe shape of one fault, no run namedLoading the four recorded runs.
Ask for a hint
Rungs open as you run the session, not as time passes. You have run it 0 times.
1. A question: available after 3 evaluated runs.
2. A narrowing: available after 5 evaluated runs.
3. A worked fragment, on two runs from a different task: available after 7 evaluated runs.
4. The answer: available after 9 evaluated runs.
04 · Compare
05 · Check your understanding
The gate: four of four, named and cited
not yetName each run's fault and cite the one diagnostic that rules out its nearest alternative. For the two runs that share a curve shape, the weight histogram or the configuration panel is what separates them. Nothing is revealed until all four go in together.
Run A
Run B
Run C
Run D
All four runs need a fault and a diagnostic before this can go in.
These are four entries from a catalogue of eight. Real runs often carry more than one at a time.
4 runs, eight possible faults. An incomplete citation can be corrected without renaming the fault.
Save your takeaway
Your takeaway
Edit it into your own words, or keep the line as it is.