Teaching code

4S010686

Teacher

Gloria Menegaz

Coordinator

Gloria Menegaz

Credits

6

Language

English

Scientific Disciplinary Sector (SSD)

ING-INF/06 - ELECTRONIC AND INFORMATICS BIOENGINEERING

Period

Semester 1  dal Oct 1, 2024 al Jan 31, 2025.

Courses Single

Authorized

Lessons timetable Moodle Seminars 0

Learning objectives

The course aims at providing competence about analysis, modeling and interpretation of multidimensional signals and images with focus on artificial vision and machine learning aspects, targeting applications in the field of multimedia and interpretable machine learning. At the end of the course the students will be able to autonomously solve typical problems requiring multidimensional signal modeling, feature extraction, analysis and interpretation of the outcomes of machine learning algorithms in the field of multimedia and artificial vision.

Examination methods

To pass the exam, students must demonstrate:
- to have understood the principles underlying visual intelligence
- to be able to present arguments on the topics of the course in a precise and organic way without digressions
- to know how to apply the acquired knowledge to solve application problems presented in the form of exercises, questions and projects.

Prerequisites and basic notions

None

Program

The course consists of two blocks: 1) signal representation using multiresolution analysis and 2) eXplainable AI methods with particular emphasis on deep learning and convolutional networks (CNN). Both blocks include a theory part and a laboratory part, which are integrated at the end of the course to form the basis for the exam project.
Part 1: Multiresolution analysis – 12 hours theory, 8 hours lab
- Revisiting the Fourier transform in 1D and 2D
- Windowed Fourier Transform
- Wavelets and multiresolution representations
- Wavelet bases Families of multiscale transforms and their properties
- Fast implementation of the transform discrete wavelet (DWT)
- Discrete wavelet transform in 2D
- Scattering transform
Part 2: eXplainabe AI – 12 hours theory, 16 hours lab
Interpretable models (White box)
- Linear regression, logistic regression, GLM, decision trees….
- Global model agnostic methods
- Partial Dependency Plots (PDP), Accumulated Local Effects (ALE), Feature Engineering
Model Agnostic Local Methods
- Local Surrogate Model (LIME), Shapley Adaptive Explanations (SHAP)
Focus on Deep Neural Networks
- Deep learning reminders
- Visualization methods (layers, filters, activation maps)
- Pixel attribution values and saliency maps
- Gradient-based methods (integrated gradients and variants)
- Layerwise Relevance Propagation
Validation of XAI results
- Associon Studies
- Proxies
Each lesson will be integrated by a practical session during the Laboratory.

Bibliography

Didactic methods

Lessons will be delivered in person. Where possible, recordings of lessons from previous years will be made available.

Learning assessment procedures

The exam involves the development and discussion of a project related to the topics covered in the course. The exam consists in the presentation of the project through slides and the discussion of the related theoretical and applicative aspects in the form of an oral interview.

Evaluation criteria

To pass the exam the student must demonstrate: - Having understood the fundamental theoretical aspects relating to the two parts into which the teaching is divided - Having understood the relationships between the topics covered - Having acquired theoretical and practical skills relating to the theory of multiresolution and its implications in the field of interpretability of deep machine learning models - Being able to transpose the skills acquired into solutions to concrete problems in a multidisciplinary field.

Criteria for the composition of the final grade

The final grade will be determined by the outcome of the presentation of the project and the discussion.

Exam language

Inglese