The course starts with an overview of canonical machine learning (ML) applications and problems, learning scenarios, etc. and then introduces foundations of Bayesian approach to solve these problems. Bayesian approach allows one to take into account subject domain knowledge and/or user’s preferences through a prior distribution when constructing the model. Besides, it offers an efficient framework for model selection. We discuss which prior distributions types are usually used, limit properties of a posterior distribution, and provide some illustrations of the Bayesian approach.

The practical applicability of Bayesian methods in the last 20 years has been greatly enhanced through the development of a range of approximate inference algorithms such as variational Bayes and expectation propagation, as well as posterior simulation methods based on the Markov chain Monte Carlo approach. As a result Bayesian methods have grown from a specialist niche to become mainstream. Therefore, we devoted a second part of the course to approximation tools, vitally important for Bayesian inference, and provide examples how to use Bayesian approaches to automatically select features, tune the regularization parameter in regression and classification, etc.

The last part of the course is devoted to advanced Bayesian methods, namely, Gaussian Processes and deep Bayesian neural networks, which have become widespread in the last 5-8 years. We discuss deep Bayesian framework and then illustrate its applications through construction of deep variational autoencoders, approaches to variational dropout, Wasserstein Generative Adversarial Networks, deep Kalman filter, etc. Home assignments include solution of applied problems, development of modifications of Bayesian ML algorithms, and some theoretical exercises.

To make the material self-contained, we also address the concept of tensorization, which allows for the creation of very high-order tensors from lower-order structured datasets represented by vectors or matrices. Then, in order to combat the curse of dimensionality and possibly obtain linear or even sub-linear complexity of storage and computation, we address super-compression of tensor data through low-rank tensor decompositions. We also explain the concepts of low-rank matrix/tensor approximations and the associated machine learning algorithms. We then elucidate how these concepts can be used to convert otherwise intractable huge-scale optimization problems into a set of much smaller linked and/or distributed sub-problems of affordable size and complexity. In doing so, we highlight the ability of tensor decompositions to account for the couplings between the multiple variables, and for multimodal, incomplete and noisy data.

In this course, we dive into RL with the goal of understanding and trying out the key principles thereof. We will study how agents interact with the environment and optimize their actions to improve rewards. Speaking of examples, imagine a video game speed run. An agent, the protagonist, interacts with the game environment and wants to beat the game as fast as possible, by dynamically adjusting his or her controls while learning on-the-fly.

At first, we will make a soft, less formal, introduction into RL and then proceed to its fundamentals – dynamical systems, agents, actors, critics etc. Secondly, we will address how RL arises from a general setting of optimal control followed by a menu of concrete methods. There is a huge amount of those and you will have the opportunity to pick one and try it our yourself. Last part of the course discusses some advanced topics of RL – convergence issues, relation to various control methods, safety.

The course comprises of 6 lectures, 6 seminars, 3 home works, 2 labs and a final project.
You will be provided with an RL framework in Python and MATLAB to play around with (the choice of the language up to you). Finally, teams of max. 5 students will conduct a research study in RL containing: an implementation of the environment, the agent, simulation studies, results evaluation. After successful completion of the home works, this final project will be rather entertaining.

The PhD students will have one extra home work assignment.

The course will be useful to all students willing to improve their understanding of natural (e.g. mineral oil) and man-made colloids and polymers. The course will also provide a hands-on experience in numerical methods of soft matter modeling via theoretical statistical approaches and mesoscale simulations, gained via homeworks and term projects. Term project will involve both physics-driven and data-driven approaches, as the students desire.

Prerequisites: Students should be comfortable with undergraduate physics, phenomenological thermodynamics and have a solid knowledge of basic undergraduate math (namely, differential equations and linear algebra, although the course requires no math beyond standard undergraduate courses offered for chemistry, physics, and engineering majors). Reasonable proficiency with any programming language is required; Matlab or Python very much desirable. Courses in machine learning required for data-driven course projects (involvement of data science is, of course, optional).

Over multiple weeks, we will investigate how researchers and practitioners use these methods and algorithms for analyzing time-series data, text data, or medical sequences.
We explore how experts use these concepts for time series forecasting, failure detection of machines, assessing the similarity of CRISP sequences, and more.

The course aims to bring all students on the same page. They do not require severe background knowledge. The objective is to provide them with both depth and breadth knowledge of the state-of-the-art in sequence modeling.

This is an introductory NLP course dedicated to classic algorithms based on the Jurafsky and Martin textbook. Later, we will offer also an additional course covering neural methods for NLP based on the Goldberg textbook.

The first part of the course provides an introduction to the methods of description, types and generation of randomness. The main idea is to establish a firm ground for more advanced topics and to help students feel comfortable with advanced machine learning courses. A special emphasis is put on the ubiquitous scale-free and non-gaussian stochastic processes also known as anomalous diffusion. Different causes of these processes will be discussed with examples such as practically important class of first-passage problems.

The second half of the course will deal with Monte-Carlo algorithms, inference and learning, classical random network theory, Markov decision processes and stochastic optimal control.

Course structure: lectures, seminars, exam.

The course is offered primarily for CDISE students who are assumed to know the basics of image processing and machine learning. In addition to the lectures, there will be 6 adapted seminars for those students who encounter the technical aspects of this course for the first time (e.g., Life Sciences students). The technical sessions will be shuffled with invited seminars by doctors/biologists with whom there is an ongoing collaboration.

Many machine learning problems are fundamentally geometric in nature. The general goal of machine learning is to extract previously unknown information from data, which is reflected in the structure (underlying geometry) of the data. Thus, understanding the shape of the data plays an important role in modern learning theory and data analytics. Real-world data obtained from natural sources are usually non- uniform and concentrate along lower dimensional structures, and geometrical methods allow discovering the shape of these structures from given data.

Originally being part of dimensionality reduction research, geometrical methods in machine learning has now become the central methodology for uncovering the semantics of information from the data.

The aim of the course is to explain basic ideas and results in using the modern geometrical methods for solving main machine learning problems such as classification, regression, dimensionality reduction, representation learning, clustering, etc.

A large part of the course addresses to most popular geometrical model of high-dimensional data called manifold model and introduces modern manifold learning methods. Necessary short information on differential geometry and topology will be given in the course.

The course lets students to be involved in meaningful real-life machine learning projects, such as mobile robot navigation, neuroimaging, to cope with challenging problems.

During this course students will be introduced with open problems in chemoinformatics and with state-of-the-art machine learning methods attempt to solve these problems. 
Particularly students will practice deep learning, including 3D convolutional neural networks and generative adversarial networks, for drug discovery and molecular design problems. The course includes three theoretical lectures, that cover basics of molecular structures, such that no prior knowledge of structural chemistry or biology is required. Seminars are python coding sessions, where students apply machine learning pipelines to derive prediction models. The end of the course comprises final project aimed to solve chemoinformatics problem of choice using machine learning.

"Lectures" in the schedule refer to approximately 3h "windows" they will are not necessary scheduled as such, but will spread out and intermitted by seminars / recitations / HW and project discussions