Artificial Intelligence

1) Knowledge and comprehension: knowledge of the classic Artificial Intelligence (AI) paradigm and of the Machine Learning paradigm for intelligent systems development, understanding of their strengths and weaknesses; knowledge of the operating logic of the main machine learning algorithms (supervised and unsupervised), of deep neural networks and of the recent neural architectures on which the current "Large Language Models" are based.

2) Ability to apply knowledge and comprehension: students will be able to autonomously prototype and validate systems that learn from data, with particular reference to textual data and Natural Language Processing (NLP) and Natural Language Understanding (NLU) tasks. They will be able to carry out the training of a model from scratch, the "fine-tuning" of a pre-trained model, choosing the appropriate metrics for performance evaluation, according to the task to be learned.

3) Autonomy of judgment: students will be able to autonomously evaluate which problems can be solved with machine learning techniques based on the available data; select the most appropriate ones among the available models and architectures for solving the task, and critically interpret the results obtained.

4) Communicative abilities: students will be able to document the problem faced, the characteristics of the dataset, the configuration of the model, and the results obtained, clearly and with proper language. They will also be able to argue their design choices.

5) Learning abilities: starting from the notions learned in the course, students will be able to understand new applications of AI by reading the reference literature in the Digital Humanities sector and to independently try their hand at learning new models and new software libraries

“Classic” Artificial Intelligence: Definition of intelligent autonomous agent in relation to the environment; Architecture of a cognitive system (perception, memory, reasoning, action, metacognition); Knowledge representation techniques; Types of reasoning: deductive and inductive, analogical, case-based, probabilistic. Uncertainty and inference.
Machine learning: Classification, regression and prediction. Supervised learning (Linear classifiers, K-Nearest Neighbours). Unsupervised learning (clustering). The "deep learning": Convolutional neural networks, outline of recurrent neural networks. Encoder-decoder architectures: the Transformers. BERT-based, GPT-based and BART-based models. Methodology for model creation, training and testing (overfitting and underfitting, dataset split in training, validation and testing, Loss functions, Cross-entropy, Softmax, model evaluation metrics; cross-validation). Self-supervised models. Fine-tuning of pre-trained models. Zero-shot, One-shot, Few-shot learning. Prompt Engineering.
The PyTORCH library for deep learning and the Hugging Face environment. Limits of models (inspectability, comprehensibility, bias in datasets)
Applications: NLP and NLU, Recommender system, Sentiment analysis, Conversational agents (Chatbot)

For “classic” AI:

Selected chapters from S. Russell, P. Norvig, Artificial intelligence: a modern approach, Upper Saddle River, Pearson, 2020 (4th edition), chapter 1 (pp. 29-39, 63-93), chapter 2 (pp. 95-140), chapter  3(pp. 147-179, 189-199), chapter 24 (pp. 856-878).

For machine learning:

- A.W. Trask, Grokking Deep Learning, New York, Manning Publications, 2019 (pp 290)

- L. Tunstall, et alii, Natural Language Processing with Transformers: Building Language Applications with Hugging Face, O’Reilly Media, 2022 (pp. 370)

Technical documentation on the machine learning libraries used for the project,  available at HuggingFace site (https://huggingface.co/docs)

Please remember that in compliance with art 171 L22.04.1941, n. 633 and its amendments, it is illegal to copy entire books or journals, only 15% of their content can be copied.

For further information on sanctions and regulations concerning photocopying please refer to the regulations on copyright (Linee Guida sulla Gestione dei Diritti d’Autore) provided by AIDRO - Associazione Italiana per i Diritti di Riproduzione delle opere dell’ingegno (the Italian Association on Copyright).

All the books listed in the programs can be consulted in the Library.