Q1. Scope, Advantages & Limitations of Machine Learning
Exam Answer
Scope: Applies to vision, NLP, healthcare, finance, recommender systems, robotics.
- Advantages: learns complex/non-linear patterns; scales with data; enables personalization & automation.
- Limitations: data quality dependence; bias/overfitting risk; interpretability issues; computational cost; model drift.
Q2. Types of Learning: Supervised, Unsupervised & Reinforcement
Exam Answer
| Aspect | Supervised | Unsupervised | Reinforcement |
|---|---|---|---|
| Signal | Label y | None | Reward r |
| Goal | Predict outputs | Discover structure | Maximize return |
| Tasks | Classification, Regression | Clustering, PCA | Policy / control |
| Example | Spam filter | K-Means | Game agent |
Semi‑Supervised: few labeled + many unlabeled. Self‑Supervised: proxy tasks (e.g. mask prediction) to learn representations.
Q3. Principal Component Analysis (PCA)
Exam Answer
Finds orthogonal axes capturing maximum variance; project onto top k components for compression & noise reduction.
- Standardize
- Covariance Σ
- Eigen decomposition
- Sort eigenvalues & choose k
- Project Z = X V_k
Q4. Logistic Regression & Decision Boundary
Exam Answer
P(y=1|x)=σ(w·x+b); boundary at w·x+b=0; train with log‑loss & gradient descent; extend to multi‑class via One‑vs‑Rest or Softmax.
- Regularization: L1/L2
- Pros: probabilistic, interpretable
- Con: linear boundary
Q5. Decision Tree: Working & Splitting Criteria
Exam Answer
Recursively split on feature with largest impurity reduction (Entropy/Gini); stop on purity or limits; leaves store class/value.
Q6. Random Forest Algorithm
Exam Answer
Bagging ensemble of trees + feature randomness reduces variance; vote/average predictions; OOB samples for validation.
Q7. Support Vector Machine (SVM)
Exam Answer
Maximizes margin; soft margin uses slack ξ with penalty C; kernels (RBF, polynomial) enable non‑linear separation.
min ½||w||² + C Σ ξᵢ s.t. yᵢ(w·xᵢ+b) ≥ 1−ξᵢ, ξᵢ≥0
Q8. Neural Network Architecture & Training
Exam Answer
Forward pass computes activations; backprop gets gradients; optimizer (SGD/Adam) updates weights; regularization (dropout, weight decay, batch norm) improves generalization.
Q9. Advantages & Disadvantages of Neural Networks
Exam Answer
Pros: Powerful non‑linear modeling; feature learning; multi‑domain success. Cons: Data/compute heavy; opaque; overfitting risk.
Q10. Guidelines for Designing ML Experiments
Exam Answer
- Define objective & metric.
- Proper train/val/test splits (no leakage).
- Baseline first.
- Pipeline fit on train only.
- Hyperparameter search.
- Fix seeds & log configs.
- Report mean±std.
- Fairness checks.
Q11. Cross-Validation Types & Usage
Exam Answer
- k-Fold / Stratified
- LOOCV (low bias, high cost)
- Time Series (expanding window)
- Nested CV for hyperparameter tuning
Q12. Clustering Approaches
Exam Answer
Partition (K-Means), hierarchical, density (DBSCAN), probabilistic (GMM).
K-Means Steps
- Init centroids
- Assign points
- Recompute means
- Repeat