	Summary
	Overview of the course
	Representing distributions as graphical models Representation: directed graphical models (Chapter 3 of KF) Directed graphical models: overview, representation of probability distribution and conditional independence statements. Representation: Undirected graphical models (Chapter 4 of KF) Representation: potentials, conditional independence and graph separability, factorization. Constructing undirected models from distributions Relationship between directed and undirected models. Common undirected graphical models: Factor models, Ising and Potts model, Gibbs distribution, log-linear models, CRFs. Feature-based potentials for flexible deployment in many applications. Application in vision and text mining.
	Inference in graphical models Overview (Chapter 9.1 of KF) Variable elimination (Chapter 9.2, 9.3 of KF) Junction trees and sum product message passing (Chapters 10.1, 10.2, 10.4 of KF)
	Learning graphical model parameters (probabilistic methods) (Lecture slides) Learning conditional graphical models (CRFs), conditional likelihood training. (Chapter 20.3.1--20.3.2) Learning with partially observed data (Chapter 19.2.2 to 19.2.2.5(inclusive) in KF)
	Sampling Foreward sampling (chapter 12.1) Importance sampling (12.2 upto 12.2.3.1) MCMC sampling (chapter 12.3): Gibbs and Langevin
	Deep Latent Representation Models for High dimensional Objects Variational Auto Encoders (VAEs) paper, Slides 1 above GANs Normalizing flows Diffusion models
	Deep models for representing sequences Transformer models Analyzing the representational power of pre-trained transformers State space models
	Inference time algorithms for Large Language Models Viewing the auto-regressive generation of an LLM as a Bayesian network, we will discuss pratical challenges of various inference algorithms. Topics covered: Limitations of greedy decoding Sampling multiple generations Grammar constrained decoding Speculative decoding Reading material
	Models for continuous variables Gaussian Process (Chapter 15--15.2.3 of https://www.cs.ubc.ca/~murphyk/MLbook/) a demo Tutorial on Bayesian optimization Time series forecasting: encoder-decoder approach as in DeepAR Correlated time series: High-dimensional multivariate forecasting with low-rank Gaussian Copula Processes, NeurIPs 2019.
	Causal Inference, Counterfactual reasoning, Attribution Understanding causality, causal graphs, identifiability Estimation from observational data Counterfactual explanations Paper: Counterfactual Invariance to Spurious Correlations in Text Classification. NeurIPS 2021
	Additional Topics State space models Parameter efficient fine-tuning In-context learning