The aim of the course is to provide students with the ability to understand the functioning of applications based on machine learning, with a specific focus on deep neural networks, as well as familiarize with the main libraries for developing them. Moreover, students will learn how to analyze, design and evaluate various solutions targeted to computer vision and, more in general, to the processing of complex data like, e.g., signals, audio and video. Specifically, students will achieve basic knowledge pertaining: - theoretical fundamentals of artificial neural networks and training algorithms (back-propagation); - main families of deep networks architectures; - hardware and software stack for deep learning; - applications of machine and deep learning for computer vision; - applications of machine and deep learning for the processing and generation of multimedia content. At the end of the course, students will be able to: - design a deep neural network, selecting the most appropriate architecture and dataset for the problem at hand; - evaluate and integrate machine learning components within multimedia applications; - implement a deep neural network using state-of-the-art libraries; - select the optimal hyper-parameters for training a neural network; - solve practical problems during development of a machine learning component, such as under- and over-fitting.

Fundamentals of computer programming. Mathematical analysis I. Linear algebra.

The course program will be articulated in in-class theoretical lessons and lab exercises, as reported below. Course introduction Fundamentals of machine learning (in-class lessons 0.9 CFU) - Basics of probability - Shallow and deep neural networks - Statistical machine learning (supervised learning, overfitting, regularization) - Back-propagation and stochastic gradient descent Methods for the implementation of deep networks (lab 1 CFU) - Automatic differentiation, programming in Pytorch - Training deep networks: hyperparameter tuning, data augmentation, batch normalization, and transfer learning Applications for computer vision (in-class lessons 1 CFU + lab exercises 0.6 CFU) - Overview and taxonomy of the main applications - Convolutional neural networks (definition and main architectures) - Neural networks for object detection and tracking - Neural networks for image segmentation (encoder-decoder) Applications for multimedia data processing (in-class lessons 1 CFU + lab exercises 0.3 CFU) - Recurrent neural networks and transformers - Audio/Speech processing: audio identification/fingerprinting and sound classification - Video analysis and processing: image/video recognition/labeling - Human movement analysis: pose estimation and action recognition Advanced topics (in-class lessons or seminars 0.8 CFU) - Introduction to generative models and GANs Project work (lab 0.4 CFU)

The course will be taught in English.

The course will encompass approximately 37 hours of in-class lessons and 23 hours of lab exercises. Exercises will be aimed to introduce the software stack used for the implementation of neural networks and the practical aspects associated with their training. During lab activities, students will be provided with problems to solve individually on in a pair, which will be followed by a discussion of the main approaches adopted to reach the solution. These activities will be preparatory to the development of an individual or group project, which will concur to the determination of the final grade. Seminars could be organized to tackle advanced topics.

Teaching material will be provided during the course and includes: - slides for in-class lessons, notes for lab exercises, sample exams and other content on the Portale della Didattica; - recordings of in-class lessons (and of lab exercises solutions); - selected tutorials and reading material. Books: - I. Goodfellow, Y. Bengio, A. Courville. Deep Learning. MIT press - Eli Stevens, Luca Antiga, and Thomas Viehmann, Deep Learning with Pytorch - Simone Scardapane, Alice in a differentiable wonderland

Lecture slides; Lecture notes; Text book; Exercises; Lab exercises; Video lectures (current year);

You can take this exam before attending the course

Exam: Written test; Compulsory oral exam; Individual project; Group project;

The exam consists of a written test and a group project, which contribute to the determination of the final grade in the measure of 1/3 and 2/3. The exam is passed if the total evaluation obtained in the two parts is at least 18/30. The written test, lasting approximately one and a half hour, is aimed at evaluating the knowledge of the topics listed in the official program of the course and will include open-ended questions or short exercises aimed at evaluating the ability to apply theoretical concepts. During the written test it is not allowed to keep and consult books, notes, sheets with exercises, forms, and similar material. It may be allowed to keep a calculator if it is functional to the exercise. The laboratory project, carried out individually or in groups of at most 3 students, is aimed at applying the notions acquired during the lessons and labs to a representative use case for the orientations involved. The project will be chosen by the group from a series of proposed topics, related to the techniques and applications addressed in the course; projects proposed by the students may be accepted pending instructors’ approval. In any case, the project shall include the implementation and training of a deep neural network. The project can consist, for instance, in the implementation of an article of scientific literature, the creation of an original application, or the resolution of a challenge. Using publicly available components or code (e.g., pre-trained models) shall be pre-agreed with the instructor. The project will define an intermediate milestone that must be completed to access the discussion phase, according to procedures and deadlines that will be communicated during the course. The project must contain the final code, at least one trained model and a short paper (6-8 pages) illustrating the techniques adopted, the values of the hyper-parameters, the dataset, the evaluation methodologies, and the results obtained. The project must be delivered one week before the discussion, which will constitute the oral exam. During the discussion, questions will ascertain the individual contribution in the implementation of the project; a demo to verify the actual functioning of the code that has been delivered may be required. Up to 2 additional points can be awarded for the lab project/oral exam (for the use of techniques and tools not illustrated by the teacher, or characterized by significant complexity, or for the completeness of the experimental tests performed), or for extra activities pertaining to the course’s content organized in the context of the Master's Degree. These points concur to obtain 30/30 with honors grade. The results achieved in the two parts and the overall evaluation are communicated through the Teaching Portal.

In addition to the message sent by the online system, students with disabilities or Specific Learning Disorders (SLD) are invited to directly inform the professor in charge of the course about the special arrangements for the exam that have been agreed with the Special Needs Unit. The professor has to be informed at least one week before the beginning of the examination session in order to provide students with the most suitable arrangements for each specific type of exam.