Related Neural Networks Links
Learn Practice Exercises Neural Networks Tutorial, validate concepts with Practice Exercises Neural Networks MCQ Questions, and prepare interviews through Practice Exercises Neural Networks Interview Questions and Answers.
Neural Networks
15 Essential Q&A
Interview Prep
Neural Network Practice Exercises — 15 Interview Questions
How to structure drills, attack MCQs, trace shapes, whiteboard backprop, and use spaced repetition before interviews.
Colored left borders per card; green / amber / red difficulty chips.
Drills
Shapes
Whiteboard
Timed
1 Effective NN practice routine.Easy
Answer: Short daily blocks: concepts (cards), numeric toy examples, one coding micro-task—repeat weak tags weekly.
2 MCQ strategy under time pressure.Easy
Answer: Eliminate impossible options (wrong units, violates invariances); for “which is true†questions, check edge cases (zeros, imbalanced classes).
3 Conv output shape (no code).Medium
Answer: Spatial: floor((W−K+2P)/S)+1 per dimension (with correct dilation if asked). Channels change via filter count, not kernel size alone.
Out = ⌊(W − K + 2P) / S⌋ + 1 (per spatial dim)
4 Count parameters in a linear layer.Easy
Answer: in×out + out for weights plus bias (if bias enabled)—common quick sanity check in screens.
5 Whiteboard: derivative of MSE w.r.t. prediction.Medium
Answer: For ½(ŷ−y)², dL/dŷ = (ŷ−y). Shows you remember scaling constants matter for manual derivations.
6 Sigmoid saturation—interview angle.Easy
Answer: Gradients ≈0 at tails → slow learning; prefer ReLU family in hidden layers; sigmoid often for binary output probability.
7 Softmax + CE gradient pattern.Medium
Answer: With CE on logits, gradient w.r.t. logits simplifies to p − y (one-hot y)—elegant result worth memorizing for speed.
8 BN during train vs eval (drill).Medium
Answer: Train uses batch stats; eval uses running mean/var learned during training—say why eval mode matters for fair metrics.
9 Dropout train vs inference.Easy
Answer: Randomly zero activations in training; at inference activations scaled (or inverted dropout in training) so expected scale matches.
10 SGD vs Adam—when prefer SGD?Medium
Answer: With careful LR schedule + momentum, SGD can generalize slightly better on some vision tasks; Adam faster early convergence—trade-off question.
11 LR too high / too low symptoms.Easy
Answer: Too high: loss spikes, NaNs. Too low: barely moves, underfits slowly—mention LR finder or grid search as practical response.
12 Spot overfitting from curves.Easy
Answer: Train metric improves while validation worsens or plateaus—response: regularization, more data, early stopping, simpler model.
13 10-minute coding drill example.Medium
Answer: Implement softmax numerically stable, or a single linear layer + CE forward on random tensors—tests API fluency without full CNN.
14 Spaced repetition for ML theory.Easy
Answer: Revisit cards at increasing intervals; tag errors (“vanishing gradâ€, “AUCâ€) and drill those stacks before interviews.
15 Mock interview structure.Easy
Answer: 5 min warm-up concepts, 20 min mixed questions, 15 min coding—record yourself; review filler words and unclear explanations only.
Keep a mistake log—interviewers love “here’s what I got wrong once.â€
Quick review checklist
- Shapes, param counts, MSE/softmax+CE gradients.
- BN/Dropout modes, optim/LR stories, overfit signal.
- Timed micro-codes + spaced repetition + mocks.