Course: Trustworthy AI

Topic outline

General
Welcome to the home page of the Trustworthy AI course. The course consists of short chapters that we will cover in each lecture. Before each lesson, please review the reading material and the videos. Each lesson finishes with homework, where you will complete the quiz. To complete the quiz, you will most often need to do a small project, where you will analyze a dataset and answer the quiz's questions with the results of your analysis.

The Trustworthy AI course is composed of two main parts: the theoretical part and the practical part. Below, a summarized description of the main contents are reported.

THEORETICAL PART (Lecturers: Valentina Beretta and Chiara Demartini) - This section is mainly dedicated to the discussion of the foundations of Trustworthy AI, the presentation of related European Framework and best practices, such as challenges in data management.

This part will provide also a focus on "Realizing Trustworthy AI solutions for diagnosis and prognosis support" (Lecturer: Alberto Signoroni). The topics discussed are the following:

• Introduction and taxonomy of learning-based systems (machine learning and deep learning) for biomedical signals and diagnostic images. Development phases of an AI system to support the diagnostic activity and role of the clinician (design, development, validation, integration, deployment, maintenance/monitoring).

• Data acquisition, selection and "curation" (formats, anonymization, annotation, normalization).

• Phases and methods for testing and validating an AI system to support diagnosis.

• Ethical and regulatory aspects for the use of reliable AI systems (explainable AI, trustworthy AI).

PRACTICAL PART (Lecturer: Emanuela Raffinetti; Python instructor: Alex Gramegna) - As statistics has become a valuable asset of AI, twelve sessions will be addressed to the illustration and discussion of the basic and advanced algorithms and models both on the theoretical and practical view point. Specifically, two sessions will be focused on the main statistical models; a session will be addressed to the model selection topic; a session will describe the most commonly used machine learning models; three sessions will provide an overview of the recent research developments in the domain of trustworthy AI concerning the most innovative metrics which can be exploited in order to evaluate the safety of AI methods. Case studies and dataset examples will be extensively presented to give students an appreciation for the application of AI methodologies to different settings. Two sessions will be managed by students who are required to read a sample of selected papers (reporting the recent statistical proposals in the AI domain), reflect on those and prepare a brief report on a specific issue (for instance, illustrating a case study in healthcare management which can be faced by the use of the main approaches illustrated along the papers). Finally, in order to strength the skills in interpreting the output generated by the highly complex AI methods, three practical sessions will be organized with the use of the Python software for the analysis of real data. These sessions aim at illustrating the implementation of the models and metrics presented throughout the course.

This course is offered in the xAIM's master's program. The course is co-financed by the European Commission instrument "Connecting Europe Facility (CEF) Telecom" through the project "eXplainable Artificial Intelligence in healthcare Management project" (2020-EU-IA-0098), grant agreement number INEA/CEF/ICT/A2020/ 2276680.

Copyright

This material is offered under a Creative Commons CC BY-NC-ND 4.0 license.

Software
- Python
Additional Material and Communication Channels
- Ethics guidelines for trustworthy AI
  
  https://digital-strategy.ec.europa.eu/en/library/ethics-guidelines-trustworthy-ai
  
  https://www.nist.gov/artificial-intelligence
- Identifying High-Risk Groups Using SHAP Values on Healthcare Data
  
  https://towardsdatascience.com/identifying-high-risk-groups-using-shap-values-on-healthcare-data-e3e7198f30f6
- Announcements Forum
TAI Assignment
- TAI Assignment
- TAI Practical Part - E. Raffinetti Assignment
Establishing the rules for building trustworthy AI
Seven essentials for achieving trustworthy AI Trustworthy AI should respect all applicable laws and regulations, as well as a series of requirements; specific assessment lists aim to help verify the application of each of the key requirements:
1. Human agency and oversight: AI systems should enable equitable societies by supporting human agency and fundamental rights, and not decrease, limit or misguide human autonomy
2. Robustness and safety: trustworthy AI requires algorithms to be secure, reliable and robust enough to deal with errors or inconsistencies during all life cycle phases of AI systems
3. Privacy and data governance: citizens should have full control over their own data, while data concerning them will not be used to harm or discriminate against them
4. Transparency: the traceability of AI systems should be ensure
5. Diversity, non-discrimination and fairness: AI systems should consider the whole range of human abilities, skills and requirements, and ensure accessibility
6. Societal and environmental well-being: AI systems should be used to enhance positive social change and enhance sustainability and ecological responsibility
7. Accountability: mechanisms should be put in place to ensure responsibility and accountability for AI systems and their outcomes European Commission, ref. IP/19/1893
THEORETICAL PART
PRACTICAL PART - Introduction to Statistical Learning models (Rec)
Statistical learning models must be considered as integral part of AI systems. They play a basic role for the formulation of the research question, the definition of the research design, for the interpretation of the results. The main topic of this sessions is introduce the main most commonly employed statistical learning models. Specifically, an overview of the linear and logistic regression models is provided. Linear regression models allow to detect the most impacting factors affecting the phenomenon under study, represented as a variable expressed on a continuous scale. Logistic regression models allow to detect the most impacting factors affecting the phenomenon under study, represented as a variable expressed on a binary scale (absence or presence of an attribute). The estimation procedures of both linear and logistic regression models will be widely discussed in order to allow students to acquire the necessary skills for the interpretation of this type of models.
PRACTICAL PART - Model Selection (Rec)
Several models can be built by varying the number of factors (predictors) which may be considered in order to "explain" the phenomenon under study (target variable). Different model configurations can be then compared to select the best one (i.e., the model characterized by the set of factors which actually provide a significant impact on the target variable). To choose the best model, statistical tests can be applied, to evaluate the significance of the predictors, together with specific metrics, which allow to assess the performance of the model in terms of its capability to predict the phenomenon of interest.
- Model Selection (Session 3) Folder
PRACTICAL PART - Introduction to Machine Learning Models (Rec)
Machine learning models represent a set of complex models which are characterized by a high predictive accuracy capability. In the session, the most commonly used machine learning models are presented. Specifically, the focus will be addressed to neural network and tree models.
- Neural Network and tree models (Session 4) Folder
PRACTICAL PART - Python Practice 1 (Statistical and Machine Learning Models - Model Selection) (Rec)
- Statistical and Machine Learning Models - Application in Python (Session 5) Folder
PRACTICAL PART - Explainability of Artificial Intelligence methods (LIVE)
Highly complex machine learning models are characterised by a "black-box" nature resulting in automated decision making which can lead to take wrong actions as a consequence of automatic choice. In order to monitor and reduce the risks deriving from the lack of transparency, the highly complex machine learning models have to be boosted with novel methodologies that allow to explain their predictive output. The main recent contributions to make AI methods explainable are discussed in this session.
- The Explainability principle (Session 6) - Session Scheduled on Wednesday 12 April 2023 - 6:00-8:30 p.m. The Zoom link is available in Moodle Calendar. Folder
PRACTICAL PART - Reading Scientific Paper 1
TO BE DONE AFTER SESSION 6 - Read the paper and provide a brief report illustrating a case study in healthcare management which can be faced by the use of the main approaches illustrated along the paper. The report has to be sent by e-mail (emanuela.raffinetti@unipv.it). Deadline: 10th June (see Moodle Calendar).
- First Reading - Paper: Lorenz Model Selection (Session 7) Folder
PRACTICAL PART - Accuracy of Artificial Intelligence methods (Rec)
One of the most relevant key-principle for a trustworthy AI is the model accuracy, intended as the capability of the model to reach accurate predictions. In literature, a wide set of metrics are available depending on the nature of the target variable to be predicted. In the case of a variable expressed on a continuous scale, the Root Mean Squared Error (RMSE) is typically considered, while in the case of a target variable expressed on a binary scale, the Area Under the Receiver Operating characteristic Curve (AUROC). In this section, a new measure based on the ranks, which evaluates the concordance between the ranks of the predicted values and the ranks of the actual values of a series of observations to be forecast, is discussed. The proposed measure is agnostic as it can be applied to any type of model and ordered target variable (continuous, ordinal or binary) improving the model comparison framework.
- The Accuracy principle (Session 8) Folder
PRACTICAL PART - Python Practice 2 (Explainability and Accuracy of AI) (LIVE)
- Explainability and Accuracy of AI - Application in Python (Session 9) - Session Scheduled on Friday 14 April 2023 - 6:00-8:30 p.m. The Zoom link is now available in Moodle Calendar. Folder
PRACTICAL PART - Reading Scientific Paper 2
TO BE DONE AFTER SESSION 8 - Read the paper and provide a brief report (already including the discussion on the first scientific paper in Session 7) illustrating a case study in healthcare management which can be faced by the use of the main approaches illustrated along the paper. The report has to be sent by e-mail (emanuela.raffinetti@unipv.it). Deadline: 10th June (see Moodle Calendar). Note that you have to send a UNIQUE report where, after you selected a specific case study in the healthcare area, you investigate it using the methodologies proposed in the two papers.
- Second Reading - Paper: Explainable AI methods in cyber risk management (Session 10) Folder
PRACTICAL PART - S.A.F.E. Artificial Intelligence (Rec)
- The S.A.F.E. score (Session 11) Folder
PRACTICAL PART - Python Practice 3 - S.A.F.E. Artificial Intelligence (Rec) - not held since the necessary Python packages need to be updated by the maintainers
Not available
Realizing Trustworthy AI solutions for diagnosis and prognosis support
- Realizing Trustworthy AI solutions for diagnosis and prognosis support (Part1 2024) File

Topic outline

Copyright

Software

Additional Material and Communication Channels