Extra mid-sem guide built from previous-year EC-2 regular/makeup papers, Lecture 8 tips and slide emphasis. Use it after one pass through cheatsheet and before solving papers.
Priority order
| Rank | Topic | Why high priority | Practice page |
|---|---|---|---|
| 1 | Autoencoder/CNN calculations | Repeated in previous papers and Lecture 8: activation/loss, BCE, backprop, Conv2D shapes, params, pooling effects. | Solved autoencoders |
| 2 | Normalizing flows | Flow transform, invertibility, RealNVP masks/log determinant, dequantization recur across paper and slides. | Solved flows |
| 3 | MADE / autoregressive models | Sampling order and mask diagrams are likely conceptual+diagram marks. | Solved AR |
| 4 | PCA/ICA/complexity | Short marks: complexity, ICA application, incremental PCA disadvantages. | Cheatsheet PCA |
| 5 | PixelCNN family | Comparison marks: Gated PixelCNN, PixelCNN++, PixelSNAIL. | Slides explained |
Previous-year pattern
- Q1 style: Autoencoder/CNN calculation. Usually activation/loss/backprop or transposed convolution.
- Q2 style: Probability/MLE or RealNVP transform.
- Q3 style: PCA complexity, ICA application, PixelCNN/flow conceptual subparts.
- Q4 style: MADE masks/sampling plus PixelCNN/WaveNet/PixelSNAIL comparison.
How to write answers
| Question type | Write this first | Common loss |
|---|---|---|
| Conv/AE numerical | Layer shape formula, parameter formula, then substituted numbers. | Skipping bias term or pooling shape. |
| AE activation/loss | Data type: binary/one-hot/real-valued. Then activation and loss. | Using MSE for binary without justification. |
| PCA/ICA | Objective difference: variance vs independence. | Saying ICA = PCA with different name. |
| MADE mask | Write factorization and dependency rule before drawing edges. | Allowing output \(x_i\) to see \(x_i\) or future variables. |
| Flow | Write invertible map, inverse if needed, Jacobian/log-det. | Forgetting absolute determinant or dequantization reason. |
One-day crash plan
- 60 min: cheatsheet + formula sheet.
- 90 min: solve autoencoder/CNN and flow solved-paper questions without expanding answers.
- 45 min: MADE masks, PixelCNN family comparison, Parallel WaveNet.
- 45 min: PCA complexity, ICA/CCA/LLE quick comparisons.
- 60 min: run question bank orally, then check weak spots.
Must memorize
- Conv2D params and output shape.
- Transposed convolution output shape.
- BCE and MSE choices.
- Laplace MLE: median + mean absolute deviation.
- Autoregressive factorization.
- Flow change-of-variables formula.
- RealNVP affine coupling and log-det.
- PixelCNN -> Gated PixelCNN -> PixelCNN++ -> PixelSNAIL improvement chain.
- Continuous point probability equals zero.
Exam-room rule. For every numerical, write formula before substitution. For every comparison, write one line for mechanism and one line for tradeoff.