FURTHER EDUCATIONAL ACTIVITIES

The course provides a compact introduction to Robust & Adaptive Deep Learning (RD-DL) and Generative Artificial Intelligence (Generative AI), with a focus on adversarial robustness and regularization (gradient/Jacobian, Lipschitz constraints), domain shift/adaptation, continual learning (stability–plasticity), generative models, and Large Language Models. RD-DL in non-standard geometric spaces will be introduced, and selected applications will be presented in the industrial, legal, and financial domains.

Knowledge and understanding: students will acquire the competencies needed to understand topics related to RD-DL and Generative AI; input-sensitivity analysis and robustness techniques; continual-learning scenarios and methods; the architectures of major generative models; the Transformer/LLM paradigm and associated training approaches; and key sector-specific applications.

Applying knowledge and understanding: students will be able to design AI-based training/evaluation pipelines (in Python); harden RD-DL and/or generative systems against adversarial attacks; apply domain-adaptation approaches; train and evaluate generative models; and build application pipelines that leverage LLM engines.

Making judgements: students will be able to analyze stability–plasticity trade-offs; select appropriate regularization techniques/AI defenses; choose between generative and discriminative approaches; define suitable metrics; and interpret results.

Communication skills: students will acquire the domain-specific language needed in Artificial Intelligence / Deep Learning, with particular emphasis on generative models; prepare concise reports and reproducible notebooks; and present design choices and technical results, including brief model cards and notes on limitations/risks.

Learning skills: students will be able to develop a personalized method to stay current on new architectures/techniques (critical reading of scientific papers/benchmarks) in RD-DL and Generative AI; apply best practices in reproducibility; and transfer competencies to new domains (industrial, legal, financial) through exercises on real-world problems.

Lectures are delivered in the classroom with the support of slides, which are made available to students. The slides do not replace the reference texts; in addition to facilitating comprehension of the lectures, they provide a detailed account of the material covered.

If necessary, following specific directions from the University authorities, instruction may be delivered in blended or remote mode, with the required adjustments to the procedures stated above, in order to adhere to the planned syllabus outlined here.

Prerequisites: To fully understand the course content, the following are required:

• Programming in C and Python (including familiarity with the PyTorch development environment).

• Foundational knowledge of Deep Learning / Machine Learning.

• Basic knowledge of Mathematical Analysis (Calculus) and Statistics.

1. Robust & Adaptive Deep Learning (RA-DL):

• Introduction and theoretical preliminaries;

• Input sensitivity;

• Adversarial attacks: Fast Gradient Sign Method (FGSM) / Projected Gradient Descent (PGD); other adversarial attack methods (overview);

• Lipschitz regularization (overview); stability analysis (Lipschitz/Lyapunov—overview);

• Jacobian regularization;

• Domain shift/adaptation (review).

2. Continual Learning (CL) – Stability–Plasticity issues in Robust & Adaptive Deep Learning:

• Definitions and scenarios; catastrophic forgetting;

• CL methods: regularization-based, replay, hybrid and adaptive approaches;

• Lipschitz analysis and meta-plasticity (overview);

• Bio-inspired models / neuromodulation (overview).

3. Generative AI Models:

• Autoencoders (AE): Denoising Autoencoder (DAE), Contractive Autoencoder (CAE); latent spaces and regularity constraints;

• Variational Autoencoders (VAE): Evidence Lower Bound (ELBO), Beta-Variational Autoencoder (β-VAE), Conditional Variational Autoencoder (cVAE);

• Generative Adversarial Networks (GANs);

• Diffusion Models.

4. LLM, Transformers, Embeddings & Retrieval:

• Transformer models: theoretical background, Encoder/Decoder architectures; attention mechanisms; applications;

• Large Language Models (LLMs): introduction and description; learning paradigms;

• Semantic embeddings;

• Use of LLM engines: prompting, indexing, re-ranking, Retrieval-Augmented Generation (RAG), chain-of-thought; brief notes on privacy, security, and robustness issues.

5. Robust & Adaptive Deep Learning and Generative AI: Applications:

• Knowledge Distillation (KD) for sustainable/embedded AI systems: teacher–student models, KD paradigms, hyper-complex deep learning (overview);

• RA-DL in non-conventional geometric spaces (overview);

• RA-DL and Generative AI in the industrial field;

• RA-DL and Generative AI in the legal/financial field.

1.    Slides provided by the teacher

2.    I. Goodfellow et al., Deep Learning

3.    D. Foster, Generative Deep Learning (2a edizione)

4.    R. Gupta et al, Generative AI: Techniques, Models and Applications, ISSN 2367-4512

5.    Scientific Papers;

Final examination will consist of:

1. Written test.

2. Development of a Python-based project agreed upon with the teacher.

The written test comprises five open-ended questions. Registration for the written test is mandatory.

Notes:

• The use of any hardware devices (programmable/scientific calculators, tablets, smartphones, smartwatches, mobile phones, Bluetooth earphones, etc.) and of personal books or documents is strictly prohibited during the written test. Any permitted documentation will be provided by the examination board during the exam.

• During written tests, backpacks, bags, and other containers may not be kept within reach and must be left at an appropriate distance. Bringing valuables is discouraged: the examination board will not take custody of any items and cannot be held responsible for any loss.

• To sit the exams, you must register via the SmartEdu portal. For technical issues related to registration, please contact the Teaching Office (Segreteria didattica).

• Late registrations by email are not accepted. Without registration, the exam cannot be taken or recorded.

Students with disabilities and/or specific learning disorders (DSA) must contact, well in advance of the exam date, both the examination board and the CInAP contact person for the DMI (Prof. P. Daniele) to request appropriate compensatory measures. Such measures must be certified by CInAP.

In-course assessments: Not foreseen.

If necessary, following specific instructions from the University authorities, the assessment may be conducted online (remote modality), with the necessary adaptations to the above procedures.

The tests are intended to provide an overall evaluation of the student’s preparation. The final suitability/pass decision will be based on the evaluation of both the written test and the project.

• The student shall describe the challenges posed by adversarial attacks in an RL-DL system.

• What are the key characteristics and critical issues of the GAN learning paradigm?

• The student shall describe the Retrieval-Augmented Generation (RAG) approach for LLM-based systems, identifying strengths and weaknesses.

Please note that these questions are purely indicative: the questions asked in the exam may differ, potentially significantly.