Tutorials

This session will cover Symbolic AI, which is a vast research field operating human-readable knowledge representations. In particular, we will focus on Pattern Structures formalism that provides a unified language to efficiently mine patterns in tables containing numbers, categories, ngrams, graphs, and other types of data. Such patterns can then be used for mining explainable-by-design implications, clusterings, and multiclass classifiers. The hand-on-section will explore a novel Paspailleur Python package (developed in Loria, set to release in May 2025) that can solve all aforementioned tasks for all aforementioned data.

This session introduces Graph Neural Networks (GNNs) and their innovative approach to understanding complex graphs. Dive into the core concepts—from graph representations and message passing to pooling and cutting-edge attention mechanisms. Using PyTorch Geometric, you’ll gain practical experience building, training, and optimizing GNN models on real-world datasets spanning molecular structures, social networks, and beyond.

This session introduces convolutional neural networks (CNNs) and their use in image processing. You’ll learn the basics, like convolutional layers and pooling, as well as advanced techniques such as data augmentation and transfer learning. Using TensorFlow, you’ll get hands-on experience building and training CNN models on real image datasets. Whether you’re new to CNNs or looking to improve your skills, this session will help you confidently work with image data and grow your expertise in image processing!

This session introduces meta-learning—’learning to learn’—which focuses on creating models that can quickly adapt to new tasks by leveraging prior knowledge. Participants will explore the fundamental concepts of meta-learning and examine key techniques such as optimization-based methods like MAML, metric-based approaches like Prototypical Networks, and model-based strategies like HyperNetworks. Additionally, the tutorial includes a hands-on session where attendees will use specialized libraries to implement and experiment with these meta-learning algorithms, providing practical experience in designing adaptive and efficient machine learning models.

This tutorial focuses on the principles and practices of fine-tuning Large Language Models (LLMs) on downstream tasks. The session will consist of two parts: a 50-minute theoretical introduction followed by a 50-minute hands-on workshop. Regarding theory, we will review the fundamental structure of LLMs, explore challenges in adapting these models to new domains as well as computational resources, and introduce techniques that make fine-tuning more accessible and resource-efficient. The practical part will guide participants through fine-tuning LLMs for a wide range of tasks on textual data, with an emphasis on understanding the parameters and methods that lead to effective training (LoRA adapters, quantization, etc.). By the end of the tutorial, attendees will be ready to efficiently fine-tune LLMs for their specific research or application needs.

The tutorial begins with an introduction to reinforcement learning, covering key concepts such as agents, environments, and the learning process. It then delves into the growing use of deep learning for tackling complex, high-dimensional challenges for the task. In the hands-on section, the tutorial will guide participants through setting up a training environment using OpenAI Gym to develop deep reinforcement learning agents.