Learning objectives

The objective of this course is to provide the theoretical and practical foundations regarding the analysis and interpretation of machine/deep learning algorithms applied to data and signals of different dimensionality and nature. Emphasis will be placed on the analysis of deep convolutional networks (CNNs) as a link between signal processing (filtering, rate change, non-linear operations) and deep learning, allowing to fix some key concepts for the analysis and interpretation of the behavior of the networks as highlighted by the scattering transform. Taking inspiration from this analysis, the concepts and methods acquired will be extended to other types of networks and algorithms so that at the end of the course the student will have acquired broad-spectrum expertise on the state-of-the-art methodologies for the analysis and interpretation of of ML and DL algorithms. Achieving this objective requires skills in two specific areas: 1) signal and image processing using multiscale methods and 2) explainability methods (eXplainable AI) allow the interpretation of the outcomes and behavior of machine/deep learning Algorithms.

The course therefore consists of two parts, dedicated to the two topics mentioned above:
Part 1: representation of signals and images using multiresolution analysis and Part 2: explainable artificial intelligence methods. The theoretical parts will be integrated with practical sessions in which the concepts acquired will be put into practice considering specific case studies.

At the end of the course the student will have acquired the fundamental skills regarding
explainability and interpretability and the related theoretical foundations as well as the ability to deal with specific case studies.

Part 1: Multiresolution Time/Frequency Analysis (3CFU)
Part 1 concerns the modeling and signals (1D-2D-3D) necessary to understand the operating modes of convolutional networks and, more generally, the analysis of the behavior of ML/DL algorithms from the signal and image processing perspective.

Part 2: Explainable Artificial Intelligence Methods (3CFU)
Part 2 delves into ML/DL methods to provide the tools that are needed for understanding
the “why” and “how” of their outcomes. After introducing the basic concepts, a taxonomy of existing eXplainable AI methods (intrinsic, post-hoc, model-specific, model-agnostic, local, global, etc.), of their properties (sensitivity, invariance implementation, separability, stability, completeness, correctness, compactness) and of the main types of explanations will be illustrated. The main feature ranking (SHAP, LIME) and visualization methods (feature attributions maps, LRP, GradCam) will then be analyzed in detail. Other methods may be considered, depending on the evolution of the state-of-the-art methodologies available. Each lesson will be complemented by a practical session during the Laboratory.