Course slides name Simon J.D. Prince's Understanding Deep Learning as the main textbook, Goodfellow/Bengio/Courville as reference, and Geron/Chollet as practical references. This page maps the mid-sem topics to book-level ideas without drifting outside the exam scope.
Understanding Deep Learning -- core map
For the mid-sem, the most relevant book-style idea is that modern generative models learn probability distributions over high-dimensional data. Chapter 16 of Understanding Deep Learning is directly referenced by the flow lectures. Read it as: define an invertible transformation, use change of variables, make the Jacobian tractable, then compose many simple invertible layers.
| Book idea | Course slide connection | Exam use |
|---|---|---|
| Latent variable / representation | PCA, autoencoders, flows | Explain why lower-dimensional codes matter. |
| Maximum likelihood | Likelihood-based models | Write objective and log-likelihood. |
| Autoregressive factorization | MADE, WaveNet, PixelCNN | Explain tractable likelihood vs serial sampling. |
| Normalizing flows | Lectures 6-7 | Change-of-variables, Jacobian, RealNVP. |
Hands-On ML -- Chapter 8 style reading
Lecture 2 explicitly points to Geron's dimensionality-reduction chapter. Use it for PCA, randomized PCA, incremental PCA, kernel PCA and LLE. The exam asks complexity, practical use and limitations, not deep proofs.
Deep Learning book -- reference reading
Goodfellow/Bengio/Courville is useful for autoencoders and generative-model framing. Focus on undercomplete vs overcomplete autoencoders, regularization, denoising and sparse constraints. Do not over-study VAE/GAN/diffusion for mid-sem unless your instructor confirms they are included.
Practical reading rule
- For calculations: formula sheet first.
- For intuition: slides explained and companion notes.
- For book depth: use UDL Ch16 for flows, Geron Ch8 for PCA variants, Goodfellow autoencoder sections for AE regularization.
- For previous-paper marks: solved paper beats extra reading.