Courses

Many algorithms in artificial intelligence result from suitable formulations of mathematical optimization problems. So it comes as no surprise to learn that these algorithms rely heavily on gradients. All too often it turns out that the evaluation of gradients is the cornerstone of the whole solution process. These gradients can be efficiently evaluated by a set of powerful techniques known as automatic differentiation. The backpropagation algorithm which is frequently used to train deep neural networks is a particular instance of automatic differentiation.

We consider the binary classification problem: Classify data into two classes such as yes and no, 0 and 1. In a supervised learning setting, a predictive model is created. Scattered data interpolation and approximation leads to a class of models. These can be transformed into the solution of partial differential equations. In the case of high dimensions, efficient and parallel techniques like sparse grids can be applied.

Inference is the process of applying trained machine learning models to new data. Often, this data is generated at runtime, e.g., through sensor or user input. Two key factors differentiate machine learning inference from training. First, the models typically go through an extensive compilation phase, which minimizes the response time once deployed. Second, inference often targets mobile devices operating under strict power constraints and with limited resources. This course gives an introduction to mobile inference. We will discuss state-of-the-art compilation techniques, including layer fusion and quantization. In the interactive sessions, we’ll then deploy our compiled models on a smartphone system-on-chip and learn how to make the model available through a user-facing Android app.

Explore the world of deep learning with PyTorch in this lecture with hands-on session. Learn the basics of building neural networks, data loading, training, testing, and more. Through practical examples, delve into convolutional layers, custom data loaders, and model saving/loading. Apply your knowledge to a real-world MNIST digit recognition task. Gain the essential skills to navigate PyTorch's ecosystem and build your foundation in deep learning.

To solve partial differential equations for simulating physical processes, such as mass or energy transport, various numerical approximation methods are available. Conventional approaches to training Artificial Neural Networks for simulating physical processes fail, as they do not incorporate the fundamental physical principles. Physics Informed Neural Networks (PINNs) stand out by incorporating the governing physical laws, as described by partial differential equations, during the training process. Therefore, PINNs enable an efficient solution of partial differential equations for the simulation of physical processes such as mass or energy transport.

This course is a specialized exploration of the fundamental skills required for securely logging into Linux servers using the SSH (Secure Shell) protocol. Designed for beginners and individuals seeking a foundational understanding of server access, it offers a practical and immersive learning experience. Through the course, participants will gain hands-on expertise in establishing secure connections to Linux servers through SSH, with a strong focus on both the command-line interface and the development environment provided by Visual Studio Code (VS Code). They will also discover how to run Python code, whether independently or through the Jupyter interface using conda environments.