Deep Learning (2021/2022)
Scientific Disciplinary Sector (SSD)
INF/01 - INFORMATICS
The teaching is organized as follows:
The course provides the theoretical foundations and describes advanced methodologies related to the area of deep learning. Deep learning solves machine learning and pattern recognition problems through the neural network paradigm and the numerical optimization. The course is also highly implementative, offering specific programming concepts for professional Python-based systems. Deep learning aims to build mainly nonlinear regression and classification systems based on neural networks. A neural network can be seen as a simple computational structure (multinomial logistic functions + non-linearities, intertwined together) which is enhanced by structuring itself at various levels (layers) of various types (fully connected, convolutional, recurring, and many others). Each of these structures underlies a very precise theory (for example the dropout of neural networks refers to the Bayesian approximation) which will be formally detailed by the teacher. In this way, the student will not only be a user of the discipline, but will manage it by acquiring formal critical skills. Particular attention will also be paid to the aspect of explainability, that is, all those techniques capable of communicating critical cases in which a particular neural network is unable to solve problems. The course will provide case studies on which to apply the studied techniques, to make them immediately usable in a professional context.
The course presents a series of state-of-the-art topics in the field of recognition. Deep learning will be analyzed starting from its methodological basis. Each topic will be explained through updated articles together with the lesson slides. The following books are suggested as a reference:
- Christopher M. Bishop. 2006. Pattern Recognition and Machine Learning (Information Science and Statistics). Springer-Verlag New York, Inc., Secaucus, NJ, USA.
- Ian Goodfellow, Yoshua Bengio, and Aaron Courville. Deep learning. MIT Press, 2016.
- Linear Regression, ridge, LASSO, elastic net
- Multinomial Logistic Classifier,
- Neural Networks,
- Convolutional Neural Network,
- Recurrent Neural Networks
- Long Short-Term Memory machine
- Transformer Network
The exam involves the discussion of a code project, which proposes a solution to an industrial classification problem. The final score will depend on the classification figure of merits achieved by the classifier and the theoretical motivations that prompted the student to choose a particular algorithm.