Third re-build of STA410 Statistical Computation / STA2102 Computational Techniques in Statistics
- Sampling: Inverse CDF, Rejection, and Importance Sampling
- Estimation: Monte Carlo (MC) integration, estimation error, improving efficiency, antithetic sampling and control variates (correlation)
- Markov Chain Monte Carlo (MCMC): High dimensional integration, Gibbs Sampling, Slice Sampling, Metropolis-Hastings,
PyMC
, Hamiltonian Monte Carlo (HMC) - Numerical precision and error and condition and linear algebra (floating point behaviour and SVD)
- Lecture Notebook
- No Coding Demo this week and we'll have a long lecture instead; the prerequesite reading becomes important for the end of this lecture and relevance continues into future material; what was being considered for the Coding Demo has instead just remained as part of the Homework [so the homework is a little longer in length than usual]
- Prerequesites: Linear Algebra
- Homework: Numerical Precision for Means and Variances
- Extra Reading: Analog versus Digital Arithmatic
- Linear Algebra: SVD/PCA/ICA/PRC, Condition, Regression VIFs, and Matrix Decompositions for Least Squares
- Prerequesites: Linear Algebra [Still (or now actually probably Even More) applicable compared to Last Week...]
- Lecture Notebook
- Coding Demo: Least Squares
- Homework: Randomized Linear Algebra
- Extra Coding: Gram-Schmidt and the Cholesky
- Extra Coding: More Least Squares
- Extra Reading: Computational Speed and Complexity
- Extra Reading: Matrix Condition Numbers
- Coding Challenge
- Reading Week
- Midterm
- From (Week 5) Direct Methods to Iterative Methods: Gauss-Seidel (GS), Successive Overrelaxation, Coordinate Descent (AKA Nonlinear GS), and Gradient Descent and AutoDiff
- Coding Demo: Splines, smoothing matrices (lowess/loess), generalized additive models (GAMs)
[including some extra broader contextual material on basis functions and regularization and penalty functions] - Lecture Notebook
- Homework: Gradient Descent
- Extra Reading: Line Search to find optimal step sizes and Conjugate Gradient Descent
- Extra Coding: Conjugate Gradient Descent
- Extra Reading: Function Spaces
- Extra Coding: Lagrange Polynomial Interpolation
- Coding Demo: Splines, smoothing matrices (lowess/loess), generalized additive models (GAMs)
- Optimization, Hessians and Jacobians, Gauss-Newton, Maximum Likelihood Estimation (score function, etc.) and Fisher Scoring and Newton's Method
- Lecture Notebook
- (+ iii) Coding Demo / Homework Notebook: classical optimization methods in TensorFlow
(with Nonlinear Gauss-Seidel, Gradient Descent, Gauss-Newton, Fisher Scoring, and Newton's Method) - ^
- Extra Reading: Variants on Newton's Method and Convergence Considerations
- Extra Coding: Newton's Method versus Secant, Fixed-Point Iteration, etc.
- Newton's Method Sandwich Estimators and IRLS (iteratively reweighted keast squares) (including M and Quasi-Likelihood estimation)
- Variational Inference, EM algorithm, Deep Learning (no Constrained optimization)
- Coding Challenge
- Final