These notes are meant to be as brief (and concise) as possible. They
are not full tutorials or lecture scripts.
- Bayesian Logistic Regression (aka. Gaussian Process classification).
- Describes the Bayesian version of logistic regression. Kernelization leads to equivalence to Gaussian Process classification; but also multi-class can naturally be handled.
- Some notes on gradient descent.
- (Gradient descent, monotonicity & stepsize adaptation, covariant & natural gradient, co- and contra-variance, relation to Newton step, Rprop)
- Factor graphs and belief
- (Graphical models,
probabilistic inference, message passing algorithms, loopy BP)
- Gaussian identities.
- (Normal and canonical
representation, product of Gaussians, linear transformation, marginals
& conditionals, entropy, Kullback-Leibler divergence, mixture of
- Basic 3D geometry (for robotics).
(Rotation representations, transformations (static, dynamic, affine,
contra-/co-variant), kinematic chains, Jacobian & Hessian)
- Markov Decision Processes.
(definition, Bellman optimality equation, Q-function, computing value
functions, value iteration, direct solution, policy iteration,
Q-learning, TD(lambda), eligibility traces)
- Influence Diagrams.
definition, inference methods in influence diagrams, relation to MDPs)
- Stochastic Optimal Control.
- (discrete time
formulation, linear-quadratic-Gaussian case, Riccati equations,
message passing formulation, classical cost formulation)