04. Week 0 — Daily Recall¶
Spaced practice. Answer from memory. If stuck, jump straight to the referenced section in 02_explainer.md.
Monday (after ELI5 + chapter 1)¶
- In the doctor analogy, what are the symptom list, the diagnosis, the confidence meter, the specialist committee, and the overthinking trap? (§ELI5)
- Why does the rule “if temperature > 100, sick” fail on many real patients? (§ELI5)
- Explain the 99% training accuracy to 60% production accuracy failure in plain language. (§1.1)
- What is overfitting, exactly? (§1.1)
- Why should a Lead AI Engineer ask about leakage and split design immediately? (§1.2)
Tuesday (after chapter 2)¶
- Bias vs variance — define each in one sentence. (§2.1)
- What pattern in train vs validation scores signals underfitting? (§2.1)
- What pattern signals overfitting? (§2.1)
- Draw the bias-variance curve from memory. What does the middle represent? (§2.1)
- Why can one linear boundary fail even when the task feels simple? (§2.2)
- Why does L1 regularization often set some weights exactly to zero? (§2.3)
- In one sentence, when would you prefer L2 over L1? (§2.3)
Wednesday (after chapter 3)¶
- Write the linear regression equation. What does each term mean? (§3.1)
- Write the MSE objective from memory. Why does it punish large errors strongly? (§3.1)
- Gradient descent — what is being updated, and in which direction? (§3.2)
- Do one update step:
w=5, gradient1.2, learning rate0.1. What is the new weight? (§3.2) - Logistic regression — what changes relative to linear regression? (§3.3)
- Why does logistic regression still produce a linear decision boundary? (§3.3)
- Compute
sigmoid(0)and explain why it matters. (§3.3) - Give one example where an interaction feature makes a linear model much better. (§3.4)
Thursday (after chapter 4)¶
- What kind of questions does a decision tree ask? (§4.1)
- Geometrically, what kind of boundary does a tree draw? (§4.1)
- Random forest — which problem is it mainly reducing? (§4.2)
- Why does averaging many trees help? (§4.2)
- Gradient boosting — what does each new tree learn? (§4.3)
- In one sentence, why does XGBoost often dominate tabular tasks? (§4.4)
- When should deep learning beat boosting instead? (§4.4)
Friday (after chapter 5)¶
- Why must train, validation, and test play different roles? (§5.1)
- What mistake turns the test set into a fake validation set? (§5.1)
- When do you need stratified k-fold? Group k-fold? Time-series split? (§5.1)
- Given
TP=18,FP=6,FN=12,TN=64, compute precision. (§5.2) - Using the same numbers, compute recall and F1. (§5.2)
- ROC-AUC vs PR-AUC — when does PR-AUC matter more? (§5.2, §5.4)
- What is calibration? Give the 0.9-confidence example. (§5.3)
- Why is 99% accuracy meaningless on a 1%-positive fraud problem? (§5.4)
- How does threshold choice trade off precision and recall? (§5.4)
Saturday (after chapter 6)¶
- Recreate five rows from the failure-fix chain in §6.1.
- Say all 15 key points from §6.2 in your own words.
- Pick any three interview questions from §6.3 and answer them aloud.
- Name three production lessons from §6.4 that are not purely algorithmic.
- List the five foundation assumptions from §6.5.
- Recite the bridge to
01_neural_network_primitivesfrom §6.6.
Sunday (cumulative)¶
- Explain the entire module using the doctor analogy only. (§ELI5)
- Draw the bias-variance curve, L1 diamond, and L2 circle from memory. (§2.1, §2.3)
- Compare logistic regression, random forest, and XGBoost in under 90 seconds. (§3.3, §4.2-§4.4)
- Design an evaluation plan for a churn model with monthly data. (§5.1)
- Define calibration, class imbalance, and leakage without notes. (§5.1, §5.3, §5.4)
- Re-answer the self-check questions in
01_weekly_plan.mdcold.