Teaching code

4S013522

Teacher

Pietro Sala

Coordinator

Pietro Sala

Credits

3

Also offered in courses:

• Attention Laboratory of the course Bachelor's degree in Computer Science
• Attention Laboratory of the course Master's degree in Computer Science and Engineering
• Attention Laboratory of the course Master's degree in Medical Bioinformatics
• Attention Laboratory of the course Master's degree in Artificial intelligence

Language

Italian

Scientific Disciplinary Sector (SSD)

NN - -

Period

Semester 2, Semester 1

Erasmus students

Not available

Courses Single

Authorized

Lessons timetable Moodle Seminars 0

Learning objectives

By the end of the course, students will be able to implement transformer architectures from scratch using PyTorch, gaining deep understanding of attention mechanisms and their practical application. They will develop advanced skills in creating Retrieval-Augmented Generation (RAG) systems, learning to effectively integrate information retrieval and text generation. They will master sophisticated text cleaning and preprocessing techniques, essential for preparing high-quality textual data. Additionally, they will apply chunking methods and semantic analysis for processing complex documents, using vector similarity for efficient information retrieval systems. Students will gain practical experience in creating intelligent agents using LangChain and will implement specialized attention models.

Prerequisites and basic notions

To successfully follow this course, students need a solid foundation in Python programming, including understanding of variables, functions, classes, and the use of external libraries. Students should be familiar with the Jupyter Notebooks environment and have practical experience with main Python libraries for data science such as NumPy, Pandas, and Matplotlib. From a mathematical perspective, fundamental concepts in linear algebra and basic statistics are required. It is also essential to have introductory knowledge of machine learning, understanding the distinction between supervised and unsupervised learning, along with an elementary understanding of neural networks and deep learning principles. Finally, it is important to possess basic knowledge in natural language processing, including tokenization processes and text preprocessing.

Program

The course is structured in progressive modules that guide students from theoretical understanding to practical implementation of advanced artificial intelligence systems. It begins with Deep Learning fundamentals using PyTorch, exploring natural language processing with Python. The core of the course is dedicated to implementing transformer architectures from scratch, deepening attention mechanisms and their practical applications. Students will then develop skills in Retrieval-Augmented Generation (RAG) systems, learning sophisticated text cleaning, preprocessing, and chunking techniques for complex documents. The program includes specialized modules on semantic analysis, information retrieval through vector similarity, and creating intelligent agents with LangChain.

Bibliography

Didactic methods

The course adopts an approach primarily based on frontal lessons integrated with laboratory sessions using Jupyter Notebooks. Each module combines theoretical explanations with immediate practical implementations, allowing students to experiment directly with code and visualize results in real-time. Educational resources include supplementary materials, example code, and access to libraries and frameworks in the field of artificial intelligence.

Learning assessment procedures

Learning assessment is based on a practical project and an oral examination. Students must choose one of the projects proposed by the instructor or agree on a personalized project with the instructor. The project must demonstrate practical application of the skills acquired during the course, including implementation of at least one of the studied systems (transformers, RAG, LangChain agents, etc.). During the oral examination, students will present their project and answer specific questions about the implementation, technical choices made, and underlying theoretical concepts. The exam aims to evaluate both practical implementation capabilities and theoretical understanding of the topics covered in the course.

Evaluation criteria

The evaluation is based on two main components that contribute equally to the final grade. The first criterion concerns accuracy in project development, evaluating implementation correctness, code quality, adherence to required specifications, and originality of adopted solutions. The second criterion focuses on the ability to respond during the oral examination, examining theoretical understanding of topics, ability to explain implementation choices, mastery of technical language, and skill in connecting theory and practice. The ability to critically analyze obtained results is also evaluated.

Criteria for the composition of the final grade

The final grade is composed of 50% accuracy in project development and 50% ability to respond during the oral examination on what was implemented in the project. Project evaluation considers technical correctness, solution elegance, implementation completeness. Oral evaluation is based on presentation clarity, depth of theoretical understanding, ability to justify design choices.

Exam language

Italiano Inglese