Inside Bocconi’s Summer School on Recent Trends in Artificial Intelligence

The program is structured as a deliberate progression. It starts with the foundations of machine learning: how models learn from data, how generalization works, what the different learning paradigms are, and builds systematically toward the state-of-the-art architectures that dominate the field today.

This sequencing matters. Many introductory AI programs skip directly to applications, leaving students with tools they can use but do not understand. This course takes the opposite approach: the goal is that by the time students encounter large language models or diffusion systems, they already have the conceptual scaffolding to understand why those architectures are designed the way they are, what trade-offs they embody, and where their limitations lie.

The program opens with the core mechanics of deep learning: how multi-layer neural networks are trained, how the backpropagation algorithm works, and how stochastic gradient descent drives optimization. These are not treated as black boxes, they are examined with enough rigor that students can reason about them independently.

From there, the course moves into convolutional neural networks, which remain the dominant architecture for computer vision tasks. Students learn the design principles that make CNNs effective, including techniques for regularization and depth that underpin the most widely used models in production today.

A substantial portion of the program is dedicated to generative models: one of the most active and consequential areas of AI research. This block covers variational autoencoders, generative adversarial networks, and diffusion models: the family of architectures responsible for recent advances in image synthesis, audio generation, and multimodal AI. Understanding how these systems are built, not just what they produce, is increasingly a baseline expectation in technical AI roles.

The natural language processing track is similarly thorough. It covers the evolution from sequence-to-sequence models through the attention mechanism and the transformer architecture, arriving at large language models and their underlying design principles. This is the technical lineage behind systems like GPT, Claude, Gemini, and the rest of the current generation of foundation models. Students who complete this track leave with a grounded understanding of what these systems are actually doing, a significant analytical advantage over those who only interact with them as users.

The program also covers reinforcement learning, including the deep reinforcement learning methods that have driven breakthroughs in game-playing, robotics, and model alignment. A dedicated module on geometric deep learning introduces graph neural networks, an architecture that has become important across drug discovery, logistics, social network analysis, and materials science.

A defining feature of this course is that the theoretical content is paired with hands-on coding sessions throughout. Students work directly with Python and leading deep learning frameworks and libraries used by both research institutions and industry practitioners worldwide. The practical sessions include training models from scratch, working with pre-trained architectures, and applying these tools to real tasks.