GO TO A SECTION

↑ ↓ to navigate · Enter to openExplore all sections →
FDEInterviews logo
← Machine Learning & Data Science / 56
mediumGoogleDatabricksOpenAI

Explain gradient descent intuitively, and how do you decide when to stop training?

Gradient descent is walking downhill on the loss surface, one step proportional to the slope. Knowing when to stop is the part that separates people who have trained models from people who have read about them: it is the validation curve, not the training curve, that tells you.

Updated Aug 2026 · Grounded in real Forward Deployed Engineer interview loops and written to a senior-engineer editorial bar.

Gradient descent is walking downhill on the loss surface, one step proportional to the slope. Knowing when to stop is the part that separates people who have trained models from people who have read about them: it is the validation curve, not the training curve, that tells you.

20 answers per topic instead of 10, plus saved progress and bookmarks · no cardor unlock all 548 remaining answers · ₹2,000 / $25

Every answer, concept and course, every paper breakdown, all published video lessons, hands-on FDE Lab missions, Premium PDF guides and companion files, the full practice-test bank and work-sample downloads. Referral Premium excludes guide PDFs and their companion files.

6 months · One payment · No auto-renewal

Study alongside free video lessons.

Keep practising

Next in this trackBuild a 10M-sample instruction-tuning dataset from 100B web docs. Design the pipeline.PremiumNext in this trackCatch hallucinated facts in LLM meeting summaries against the transcript, on a tight human-review budget.PremiumNext in this trackRoute each request to a cheap 7B or expensive 70B model to maximize satisfaction minus cost. Design the online router.Premium
FEDITOR'S NOTE

The screen is two-part: a clean intuition for gradient descent (downhill on the loss surface, step size = learning rate, the gradient points uphill so you go opposite) AND a real stopping policy. Strong candidates stop on validation loss (early stopping with patience) rather than a fixed epoch count or training loss hitting zero, and mention plateaus, learning-rate schedules, and restoring the best checkpoint. The reserved follow-up is 'why not just train until training loss is near zero': because that is overfitting; training loss keeps falling while validation loss turns up, and the turn-up is the stop signal.

DISCUSSION · 0

No comments yet — be the first to share your approach.