KSETA Course with Prof. Dr. Ezequiel Alvarez (UNSAM): Bayesian Machine Learning for Scientific Research
Bayesian Machine Learning for Scientific Research
Prof. Dr. Ezequiel Alvarez (UNSAM)
Course Overview
In this hands-on course, you’ll learn how to build, infer, and interpret physical systems through probabilistic models using Bayesian Machine Learning. The course relies on probabilistic programming, in particular in STAN — the state-of-the-art platform for statistical modeling and high-performance computation.
- Intuitive introduction to Bayesian reasoning
- Stan language essentials & probabilistic graphical models
- Models, MCMC sampling & diagnostics
- Best practices for model evaluation and selection
- Applications in data science, forecasting, and more
- Content: 1/3 statistics, 1/3 programming, 1/3 hands-on!
- 5 lectures x 3hs
Level: Intermediate (basic statistics & basic programming assumed)
Bayesian Machine Learning
Bayesian Machine Learning consists of combining theory, software, and a series of statistical techniques developed in recent years, to a system or problem, in order to maximize understanding based on the observed data.
Within this framework, a real-world system is modeled as a probabilistic model in which the system's data are sampled from a probability density function (PDF). This task is part art and part craft, requiring not only statistical knowledge but, above all, a deep understanding of the specific problem at hand. In fact, a crucial aspect of this art and craft involves astutely modeling the system to explicitly reveal internal variables that are not directly observed (latent variables), but about which we have prior knowledge that can be exploited in the form of priors. The framework is completed by deploying the idea and theory onto suitable probabilistic software, in which the expressions are defined over random variables instead of deterministic variables. The execution of such a script relies on Bayes' theorem to return the user a posterior probability distribution over the parameters and latent variables, which are connected to the real parameters of the system under study.
In this way, modeling, implementation, and execution within the Bayesian Machine Learning framework can provide a unique understanding of the problem based on the observation of the data. This understanding can be exploited in various ways, such as measuring parameters of the problem, determining relationships and dependencies between internal parameters, identifying the composition of a data sample, detecting anomalies in the data, sampling synthetic data from the system, and in particular, gaining a deeper and more detailed understanding of the system, among other applications.
Last but not least, you can often expect Bayesian Machine Learning to outperform Neural Network analyses for several reasons. One key advantage of Bayesian ML is its ability to incorporate prior knowledge, which is extremely valuable. In Bayesian ML, you can obtain meaningful results and draw key conclusions with relatively few data points—ranging from about 50 to 10,000—a region in which Neural Networks typically do not have enough data to learn more than what we usually do.
Real World problems don't have much data and prior info is crucial.... This is why Bayesian ML provides a unique option that should be tested in most of Nature problems!
____________________________
Please register until May 27, 2025
The Course will be held at KIT Campus North, Bldg. 425, seminar room 206.
This course is part of the Double Doctoral Degree in Astrophysics between UNSAM and KIT.
If you have to cancel your participation after registration, please modify your registration or send an email to Raquel Lujan.
Please note that you must have attended the entire courseto be added to your KSETA Transcript.