FDD3601 Deep Generative Models and Synthesis 7.5 credits

• Relevant concepts from probability theory and estimation
• Introduction to synthesis tasks and generative models
• Principles of synthesis versus classification
• Regression versus probabilistic modelling
• Modelling goals and evaluation
• Mixture density networks (MDNs)
• Autoregression and large language models (LLMs)
• Normalising flows
• Variational autoencoders (VAEs)
• Diffusion models and flow matching
• Generative adversarial networks (GANs)
• Subjective evaluation
• Hybrid approaches
• Recent developments
• Ethical and societal aspects of generative AI

After passing the course, the students should be able to:

• characterise synthesis tasks, deep generative methods, and their applications
• distinguish different objectives, performance measures, and common problems with generative modelling
• describe the relation between deep generative models and regression-based methods
• train and tune deep generative models on different datasets
• evaluate generative models objectively and subjectively
• discuss ethical aspects of particular relevance to generative AI

in order to

be able to judiciously use deep generative modelling to solve problems in industry and/or academia.

• Good programming skills including Python, PyTorch, Jupyter Notebooks.
• Algebra and geometry including vectors, matrices, systems of linear equations, inner and outer products, norms, triangle inequality, metric spaces, determinants, eigenvalues, linear dependence, subspaces, trace of a matrix.
• Single-variable calculus including functions, domains, ranges, monotonicity, exponentials and logarithms, limits, sequences, change of variables, convex functions, ordinary differential equations (ODEs), Euler’s method.
• Multivariate calculus including partial derivatives, multivariate chain rule, change of variables, gradients, Hessian and Jacobian matrices, Fourier series.
• Probability theory including probability, conditional probability, Bayes’ law, independence, random variables, probability mass and density functions, samples, random sampling, expectation/mean, variance, standard deviation, median, correlation, covariance, uniform distributions, multivariate Gaussian distributions and their properties, conditional expectation, parameter estimation, maximum-likelihood estimation, consistency, change of variables, Jensen’s inequality, Markov chains, least-squares regression.
• Machine learning including optimisation, loss functions, train/val/test sets, mean squared error, classification, accuracy, overfitting, Gaussian mixture models, high-dimensional geometry/curse of dimensionality, baselines, ablation studies. Information theory for machine learning including entropy, bits, cross-entropy, Kullback-Leibler divergence.
• Deep learning including feed-forward networks, activation functions, ReLU, softmax, CNNs, RNNs, residual networks, skip connections, U-Nets, transformer architectures, self-attention, position embeddings, mean and variance normalisation, initialisation, hyperparameters, stochastic gradient descent, updates, epochs, dropout.

• LAB1 - Laboratory work, 7.5 credits, grading scale: P, F

Examination is based on several exercises which includes both programming and theoretical questions and contains different parts for higher grades.

The final grade is determined by accumulating the grades in individual exercises.

Josephine Sullivan

Gustav Henter

Hossein Azizpour

• All members of a group are responsible for the group's work.
• In any assessment, every student shall honestly disclose any help received and sources used.
• In an oral assessment, every student shall be able to present and answer questions about the entire assignment and solution.