Intellect Not-for-Profit Foundation for the Advancement of Science and Education

Initiatives
Special courses
Neural networks and their application in space research applications

ABOUT THE COURSE

ACADEMIC STAFF

ABOUT THE COURSE
ACADEMIC STAFF

The course will cover the differences between machine learning and deep learning, as well as the PyTorch framework for writing neural networks, training them, validating and testing models, creating custom dataset classes, and more.

Participants of the course will learn to independently design neural network architectures to solve applied tasks such as segmentation, detection, classification, regression, and time‑series forecasting.

The course will thoroughly address model training challenges, specifics of working with real‑world data (including space‑related data), and key modern approaches and methods for implementing, training, and working with neural network models. These will help students quickly get accustomed to and understand the functionality of AI algorithms, and ultimately learn to apply these tools in their own scientific research. This is the main goal of the course.

Topics covered in the course

- Introduction to deep learning and machine learning, and the relevance of their application in space research.

- Introduction to machine learning: main types of problems solved with ML methods, quality evaluation metrics, and validation.

- Classical machine learning: linear models, support vector machines, k‑nearest neighbors, decision trees, and random forests.

- Feature generation and selection, model ensembles. Examples of applying ML models to space‑related applied tasks (analysis of satellite images, refining satellite location data, classification of light curves).

- Neural networks: multilayer perceptron. Backpropagation algorithm. Convolutional neural networks and their application to the analysis of Earth remote sensing images. Satellite image classification task.

- Optimization of neural networks: subtleties of training and validation during model development. Model soups.

- Recurrent neural networks: RNN, LSTM, GRU blocks. Forecasting and analysis of time series. Forecasting the evolution of satellite orbits using TLE, ILRS, and ephem catalogues as examples. Semantic segmentation. The task of identifying natural and man‑made textures in images. Specifics of working with multispectral images.

- Autoencoders and representation learning. Application of autoencoders for satellite image reconstruction and detection of telemetry anomalies.

- Project consultations within the course: assistance in formulating tasks for applying AI in students’ research and course projects.

Classes are held at the Faculty of Space Research of Lomonosov Moscow State University

The course programme includes 10 lectures, 8 seminars, 2–3 project consultations, a project defence, and 7 homework assignments with prepared templates in .ipynb notebooks (Google Colab)

Course prerequisites:

- Basic proficiency in Python programming

- Basic knowledge of mathematical analysis, linear algebra, and statistics

- Basic understanding of probability theory

- Familiarity (preferred) with the Python libraries PyTorch, TensorFlow, or scikit‑learn

- A research task where the application of artificial intelligence methods is anticipated

Teachers:

Nikita V. Belyakov

Education

Fundamental Mathematics and Mechanics, Faculty of Space Research
MSU
Areas of Expertise
Remote sensing of the Earth, satellite data processing, machine and deep learning in space research

Занятия проводятся в ауд. 851 факультета космических исследований МГУ им. М. В. Ломоносова

В программе курса 10 лекций, 8 семинаров, 2-3 консультации по проектам, защита проектов, а также 7 домашних заданий с заготовленными шаблонами в .ipynb блокнотах (google collab)

Формат проведения: офлайн с записью лекционного материала

Старт курса: 2-3 учебная неделя сентября 2024, дата уточняется

ABOUT THE COURSE

ACADEMIC STAFF

Neural networks and their application in space research applications