Skip to content

ischurov/scientific-computing-2024

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

20 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Scientific Computing 2024

Scientific Computing is an extra course I teach at the JetBrains and Constructor University (Bremen) Software, Data and Technology bachelor probgram in fall semester of 2024-25 academic year. This repository will contain materials of the course (lecture notes, notebooks), available under MIT license.

The course is aimed at bridging the gap between pure mathematical courses and machine learning. We will discuss various concepts, some of which are already known to the students and some are new, in applied and computational context. We will also learn how to use various programming tools to interact with mathematical objects and visualize them.

Prerequisites

The students are expected to know Calculus (of one and many variables), Linear Algebra and be able to program in Python.

Topics

One topic is approximately one week (maybe more), 2 lessons, 75 minutes each lesson.

Overview of scientific Python ecosystem

Jupyter notebooks: pro and contra, best practices. Package managers: pip and conda. Virtual environments. Reproducibility issues. Basic numerical and data science packages: numpy, scipy, matplotlib, pandas. Introduction to numpy: array vs. list, why do we need arrays, idea of vectorized calculations. Plotting of basic graphs with matplotlib.pyplot. Interactive widgets, %matplotlib widget. Animations with IPython.display.

Numerics in Python

Numerical types

Numerical types in Python. int vs. np.int64. Floating point arithmetic. NaN and Inf. 0.1 + 0.2 != 0.3. Numerical precision. Machine epsilon. Numerical (in)stability. Example: numerical derivative. Calculation in logarithms, logsumexp, example: softmax.

More on numpy

Shape, reshaping data. Advanced indexing, np.where. Broadcasting rules. Matrix multiplication. np.linalg. How to get rid of for loops with numpy magic. Memory cost of vectorization. Processing by batch.

pandas

pd.Series and pd.DataFrame. Constructing dataframes. Indexing. Queries. Grouping and aggregating. Concatenating and merging dataframes. Melting and pivoting.

Visualization I: matplotlib and low-level stuff

Introduction to plotting in Python. Example: gradient descent.

Visualization II: seaborn, plotly and data

How to choose the best visualization method: case study.

Linear algebra I

Linear systems. Over- and underdetermined systems. Pseudoinverse, pseudosolution, normal pseudosolution. Geometrical interpretation. Matrix factorizations. Application of factorizations to solution of a linear system.

Optimization

Theory recap: optimization of function of one variable, conditional optimization, Lagrange multipliers. Numeric optimization of function of several variables. Gradient descent. Example: gradient descent near a minimum with ill-conditioned Hessian. Newton's method. Optimization of matrix-vector functions. Optimization out of the box: scipy.optim.

Probability I

Random variables. Discrete and continuous distributions. Cumulative distribution function. Expected value, variance. Generating (pseudo) random numbers. Seed. np.random and scipy.stats. Probability density function and histogram. Rejection sampling. Transformation of random variables. Entropy. KL-divergence.

Probability II

Independent and non-independent random variables. Joint distribution (continuous + continuous, continuous + discrete, discrete + discrete). Covariance and correlation. Variance-covariance matrix and correlation matrix. Conditional distribution, conditional expectation. Central limit theorem. Monte-Carlo integration. Mutual information.

Linear algebra II

More on matrix factorization. Symmetric matrices and quadratic forms. Positive definite matrices, their properties. SVD decomposition, low-rank approximations. Example: covariance matrix as a quadratic form.

Statistics I

Random variables and samples. Statistical estimates. Consistency, unbiasedness. Likelihood. MLE.

Statistics II

Statistical hypothesis testing. Type I and type II errors. Student's t-test. Permutation test.

Bonus: elements of Bayesian statistics

Bonus: analysis of graph data

Assessment

Every week, there will be a homework. Each homework consists of problems of various difficulty. It is indicated how many points the complete and correct solution of each problem earns. Some of the problems are marked as bonus. The final grade (on a scale from 0 to 10) is calculated as a sum of all earned points (including points for bonus problems) multiplied by 10 and divided by the sum of all points of non-bonus problems.

For each homework assignment, there will be specified a due date and a hard deadline. If the assignment is returned after the due date, but before the hard deadline, its score is multiplied by $e^{-t/86400}$, where $t$ is number of seconds passed since the due date. If the assignment is returned after the hard deadline, it is scored as 0.

About

Bridging the gap between mathematical courses and ML

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published