References used to build the UDL vault, plus the source files in the local udl/ folder.
Local source files
- Lecture 1: course overview, generative-model taxonomy, self-supervised learning motivation.
- Lecture 2: PCA, randomized PCA, incremental PCA, kernel PCA, ICA, CCA/KCCA, LLE.
- Lecture 3: supervised vs unsupervised, undercomplete/overcomplete/regularized/convolutional/deep autoencoders.
- Lecture 4 and 5: likelihood-based models, maximum likelihood, autoregressive models, MADE, WaveNet, PixelCNN variants.
- Lecture 6: flow foundations, change of variables, 1D and 2D flows.
- Lecture 7: MAF, IAF, RealNVP-like architectures, Glow, Flow++, FFJORD, dequantization, applications.
- Lecture 8 PPTX: mid-sem tips, assignment-style CNN/AE/flow/MADE examples.
- Solution docs: EC-2 regular/makeup previous-year solved papers used for worked Q&A.
Books
- Simon J.D. Prince, Understanding Deep Learning. Main course textbook. Chapter 16 for normalizing flows.
- Ian Goodfellow, Yoshua Bengio, Aaron Courville, Deep Learning. Reference for autoencoders, representation learning and generative-model background.
- Aurelien Geron, Hands-On Machine Learning with Scikit-Learn, Keras, and TensorFlow. Practical PCA/dimensionality-reduction reference, especially Chapter 8 style material.
- Francois Chollet, Deep Learning with Python. Practical neural-network and autoencoder implementation reference.
Papers and model families worth knowing
- MADE: Masked Autoencoder for Distribution Estimation.
- WaveNet: autoregressive audio with causal dilated convolutions.
- PixelRNN / PixelCNN / Gated PixelCNN: image autoregressive density models.
- PixelCNN++: improved PixelCNN with mixture of logistics and architecture changes.
- PixelSNAIL: masked convolution plus causal self-attention for better long-range context.
- RealNVP / Glow / Flow++ / FFJORD: normalizing flow families.
- DDPM / GAN / VAE papers: beyond mid-sem scope in this source set, but part of full UDL map.