Neural Networks and Deep Learning

IMPORTANT NOTICE:
This course will start on January 12, 2027, at 9:00.
Lectures will be opened to everybody and will be given online on the following channel.
Please, connect 10 minutes before 9:00 to avoid disturbing the lecture.

Registration

To access the channel of the lectures and receive notifications about any change on course lectures, please register here by December 15th.

Certification of attendance

If you need a certification of attendance, be aware that the number of hours that will be certified are those recorded by Teams.

Course Program and Lectures Schedule

This course includes four modules, which can also be taken independently:

Theoretical Foundations: This module (40 hours) explains many fundamental concepts of neural networks, including unsupervised learning, clustering algorithms, supervised learning, autoencoders, convolutional networks, recurrent networks, transformers, semi-supervised learning, multi-object tracking, and generative networks. It is highly recommended to attend this module, even if you are familiar with some topics, since many concepts explained here will be used in the other modules.
Responsible teacher: Giorgio Buttazzo.

Reinforcement Learning: This module (20 hours) focuses on reinforcement learning (RL), including Q-learning, deep RL, policy gradient methods, actor-critic methods, stochastic RL, curiosity-based learning, imitation learning, and model-based RL.
Responsible teacher: Giorgio Buttazzo.

Trustworthy AI and Applications: This module (20 hours) focuses on how to make deep networks more trustworthy, interpretable, and secure.
Responsible teacher: Giulio Rossolini.

Hardware Acceleration: This module (20 hours) covers practical and implementation issues and hardware acceleration methods.
Responsible teacher: Federico Aromolo.

Lecture schedule (pdf)

Part I: Theoretical Foundations (Giorgio Buttazzo)

Course program (pdf)

Basic concepts and learning paradigms
Hopfield networks
Unsupervised learning
Clustering algorithms
Supervised learning: From perceptron to Backpropagation
Supervised learning: Important Remarks
Supervised learning: Performance metrics
Supervised learning: other approaches (RBF networks, PINNs)
Towards Deep Neural Networks
Convolutional Neural Networks (CNNs)
Autoencoders
CNNs for object classification
CNNs for object detection
CNNs for image segmentation
Deformable Convolutions
Anchor-free detection networks
Recurrent neural networks
Natural language processing
Transformers
Semi-supervised learning
Neural networks for multi-object tracking
Generative networks

Part II: Reinforcement Learning (Giorgio Buttazzo)

Course program (pdf)

The reinforcement learning paradigm
Value estimation methods
Control methods: Q-learning and variants
Examples of applications
Implementing Q-learning in C
Deep Reinforcement Learning
Actor-Critic methods
Policy gradient methods
Stochastic reinforcement learning
Curiosity-based learning
Imitation learning: direct policy learning
Imitation learning: inverse reinforcement learning
Imitation learning: adversarial reinforcement learning
Model-based reinforcement learning

Part III: Trustworthy AI and Applications (Giulio Rossolini)

Please visit the following link: Part III: Trustworthy AI

Part IV: Hardware Acceleration (Federico Aromolo)

Please visit the following link: Part IV: Implementation Issues

Exam

The exam (for those who need to take it) consists in a project development. The project is one for each module and must be agreed with the teacher responsible for the module. Be aware that the number of hours that will be certified are those recorded by Teams. The completion of the exam requires the project discussion and the delivery of the project code, including a report describing the work done. Please, read carefully the project rules in the link below.

Project rules