Training and Research

Credits

3

Language

Inglese

Class attendance

Free Choice

Learning objectives

The course is designed to bring students up to speed on the mechanics of modern AI models to establish a shared baseline. From there, we will cover the fundamentals of Automatic Differentiation (AD) and rapidly progress to the mechanics of differentiating through fixed points (utilizing techniques like implicit differentiation and phantom gradients). Finally, the course will survey cutting-edge literature, exploring how to bridge the gap between fixed-computation and iterative reasoning , address induction biases in sequence models , and build introspective networks that dynamically adjust to input complexity.

Prerequisites and basic notions

NA

Program

Key Topics Covered:
• Modern AI Architectures: A rapid recap of current state-of-the-art models to get everyone up to speed.
• Advanced Automatic Differentiation: Mechanics of AD and backpropagation through iterative algorithms.
• Fixed-Point Iteration in Deep Learning: Techniques for differentiating through fixed points without unrolling, ensuring a constant memory footprint.
• Adaptive Computation & Introspection: Designing models that dynamically adjust their computational depth and allocate resources based on the difficulty of the task.
• State Tracking and Induction Bias: Analyzing the in-distribution and out-of-distribution capabilities of sequence models.
Proposed Schedule (6 Lectures):
1. Lecture 1: Thursday April 9, 2026 – 11:00
2. Lecture 2: Friday, April 10, 2026 – 11:00
3. Lecture 3: Monday, April 13, 2026 – 11:00
4. Lecture 4: Wednesday, April 15, 2026 – 11:00
5. Lecture 5: Friday, April 17, 2026 –11:00
6. Lecture 6: Monday, April 20, 2026 – 11:00
Note on Course Material: This course dives into the bleeding edge of AI research. Much of the reading material, discussions, and core concepts will be based on papers published within the last 12 months—including highly recent preprints and active conference submissions.

Didactic methods

In presence

Learning assessment procedures

Discussion on the course topics

Students with disabilities or specific learning disorders (SLD), who intend to request the adaptation of the exam, must follow the instructions given HERE

Assessment

Ability to critically present the course topics

Criteria for the composition of the final grade

The grade will be in the form Pass/Fail. The course corresponds to 3 CFU (1 CFU/4 lesson hour).