With more and more data available for analysis and decision making - from web documents and digital media to sensory data from cameras, microphones, and ubiquitous devices - it becomes increasingly more important to understand how such large volumes of data can be analyzed by computers and used as the basis for new intelligent services, for decision making, and for making computers learn from experience. In companies around the world, from retail and banks all the way to Google, intelligent learning and analysis techniques are used to improve business decisions. Likewise, in science, important discoveries are made easier by automated learning methods, and games and other artifacts are being made adaptive with learning technology.

Within the intelligent systems track of the computer science Master's program, the Algorithms for Data Science course offers in-depth knowledge of different aspects of big data analytics and systems, including algorithmic techniques for analyzing structured and unstructured data that cannot be stored in a single computer because it has enormous size and/or continuously arrives with such a high rate that requires immediate processing. In particular, the topics include classical data mining tasks for massive data and/or data streams, mining massive graphs, and similarity search in massive data.

In this seminar we study philosophical theories of scientific explanation and understanding and explore how to apply them to formal approaches to explainability for deep learning and other neural network architectures. The seminar does not presume a deep background in philosophy or technical training in AI, and is suitable for all students of computer science or philosophy who would like to learn more about how philosophical ideas relate to developments in AI.

The seminar will be held as a block seminar. Further seminar dates will be announced during the preliminary meeting.

Simply put, forecasting means to predict the future behaviour of some observable O, conditioned on some context C — where the context here is given by a sequence of past observations on O.

By zero-shot forecasting, we mean the pretraining of a deep neural network model offline, on some large and varied time series dataset, and its application on some target time series data, which is not (by any traditional standard) in-distribution, without the need of any finetuning.

In this lab, we will explore different ideas from both the time series and NLP communities, and (try to) develop simple algorithms for zero-shot forecasting. Somewhat more in detail, we will study and make us of (i) neural operators (i.e. neural networks that learn mappings between infinite dimensional spaces), (ii) Kolmogorov-Arnold networks (a recent alternative to feedforward NN), and (iii) Transformer networks to build our models.

Our main goals will be to:

• test these ideas with some coding exercises we will do from scratch
• read and discuss some current issues of the methodology and their possible solutions.

What you will actually have to do: we will ask you to give either two or three presentations about your numerical experiments, and hand in a final report. You will work in groups of up to 3 people.

Requirements: working knowledge of machine learning libraries like PyTorch or Jax.

Registration: No more spots available.