Mario VENTO | ADVANCED MACHINE LEARNING

cod. 8860500002

ADVANCED MACHINE LEARNING

	8860500002
	DEPARTMENT OF INFORMATION AND ELECTRICAL ENGINEERING AND APPLIED MATHEMATICS
	Corso di Dottorato (D.M.226/2021)
	INFORMATION ENGINEERING
	2024/2025

CURRICULUM INFORMATION TECHNOLOGIES FOR DIGITAL MEDICINE

	OBBLIGATORIO
	YEAR OF COURSE 1
	YEAR OF DIDACTIC SYSTEM 2024
	AUTUMN SEMESTER

TEACHING UNITS

	SSD	CFU	HOURS	ACTIVITY	TYPE OF ACTIVITY
	ING-INF/05	3	18	LESSONS

PROFESSORS

	ANTONIO GRECO T Curriculum
	MARIO VENTO Curriculum

	Objectives
	To understand the theoretical foundations and the expected advantages (and disadvantages) of the state-of-the-art deep learning architectures. To understand the main limitations of traditional training processes, and the techniques proposed for improving the training effectiveness. To be able to choose or design an effective architecture for a complex problem. To design a learning system for processing non-vectorial data (sequences, trees or graphs). To exploit advanced techniques for accelerating the convergence of the training process for a complex architecture, or for transferring the knowledge learned in a different domain to a new problem.

To understand the theoretical foundations and the expected advantages (and disadvantages) of the state-of-the-art deep learning architectures. To understand the main limitations of traditional training processes, and the techniques proposed for improving the training effectiveness. To be able to choose or design an effective architecture for a complex problem. To design a learning system for processing non-vectorial data (sequences, trees or graphs). To exploit advanced techniques for accelerating the convergence of the training process for a complex architecture, or for transferring the knowledge learned in a different domain to a new problem.

	Prerequisites
	The activity requires previous knowledge of machine learning. In particular, the student is expected to know the foundations of machine learning systems, artificial neural networks, multi-layer perceptrons, deep learning and recurrent neural networks. It is assumed that the student is able to use Python and Pytorch.

	Contents
	Recalls on neural networks •Linear neural networks for regression •Linear neural networks for classification •Multi-layer perceptrons (Lectures: 2 hours; Laboratories: 1 hours) Convolutional neural networks for object classification •Convolutional neural networks •Modern convolutional neural networks: AlexNet, VGGNet, GoogleNet, ResNet, DenseNet (Lectures: 2 hours; Laboratories: 1 hours) Modern CNN architectures for object detection •Two-stage and single-stage object detection •Evolution of CNNs for object detection: from R-CNN to YOLO (Lectures: 2 hours; Laboratories: 1 hours) Transformers •Limitations of RNNs •Transformer Architecture •Encoder (Self Attention, Multi-head Attention, Output) •Decoder (Masked multi-head attention, Encoder-decoder attention, Output) (Lectures: 2 hours; Laboratories: 1 hours) Self-supervised learning and LLMs •Self-supervised learning •Transformers for text representation and generation: LLMs •Encoder-only LLMs •Decoder-only LLMs •Encoder-decoder LLMs (Lectures: 2 hours; Laboratories: 1 hours) Multi-task learning •Multi-label vs Multi-task learning •Multi-task architectures •Multi-task attention networks •Multi-task loss functions •Multi-task dataset management (Lectures: 2 hours; Laboratories: 1 hours)

Contents

Recalls on neural networks
•Linear neural networks for regression
•Linear neural networks for classification
•Multi-layer perceptrons
(Lectures: 2 hours; Laboratories: 1 hours)

Convolutional neural networks for object classification
•Convolutional neural networks
•Modern convolutional neural networks: AlexNet, VGGNet, GoogleNet, ResNet, DenseNet
(Lectures: 2 hours; Laboratories: 1 hours)

Modern CNN architectures for object detection
•Two-stage and single-stage object detection
•Evolution of CNNs for object detection: from R-CNN to YOLO
(Lectures: 2 hours; Laboratories: 1 hours)

Transformers
•Limitations of RNNs
•Transformer Architecture
•Encoder (Self Attention, Multi-head Attention, Output)
•Decoder (Masked multi-head attention, Encoder-decoder attention, Output)
(Lectures: 2 hours; Laboratories: 1 hours)

Self-supervised learning and LLMs
•Self-supervised learning
•Transformers for text representation and generation: LLMs
•Encoder-only LLMs
•Decoder-only LLMs
•Encoder-decoder LLMs
(Lectures: 2 hours; Laboratories: 1 hours)

Multi-task learning
•Multi-label vs Multi-task learning
•Multi-task architectures
•Multi-task attention networks
•Multi-task loss functions
•Multi-task dataset management
(Lectures: 2 hours; Laboratories: 1 hours)

	Teaching Methods
	The course includes 12 hours of lectures and 6 hours of supervised laboratories. Lectures are aimed at acquiring knowledge about the evolution of deep neural networks, modern convolutional neural networks, transformers, self supervised learning, LLMs and multi-task learning. Laboratories are aimed at acquiring applied knowledge and skills about specific practical aspects related to the topics of the lectures. The course also includes a project work, consisting in the design, training and validation of a deep learning approach for image or text analysis, aimed at acquiring the following applied knowledge and skills: realization of neural network architectures tailored to specific tasks, optimization of the learning procedure using custom loss functions, regularization techniques, and data augmentation to improve model generalization and speed-up the convergence, employment of robust evaluation metrics, validation techniques, and performance monitoring to ensure that the model meets the requirements without overfitting or bias. Attendance to the lectures and laboratories is mandatory; in order to be admitted to the exam, the student must participate to at least 70% of the hours.

Teaching Methods

The course includes 12 hours of lectures and 6 hours of supervised laboratories.
Lectures are aimed at acquiring knowledge about the evolution of deep neural networks, modern convolutional neural networks, transformers, self supervised learning, LLMs and multi-task learning.
Laboratories are aimed at acquiring applied knowledge and skills about specific practical aspects related to the topics of the lectures.
The course also includes a project work, consisting in the design, training and validation of a deep learning approach for image or text analysis, aimed at acquiring the following applied knowledge and skills: realization of neural network architectures tailored to specific tasks, optimization of the learning procedure using custom loss functions, regularization techniques, and data augmentation to improve model generalization and speed-up the convergence, employment of robust evaluation metrics, validation techniques, and performance monitoring to ensure that the model meets the requirements without overfitting or bias.
Attendance to the lectures and laboratories is mandatory; in order to be admitted to the exam, the student must participate to at least 70% of the hours.

	Verification of learning
	The exam consists in a project and an oral interview. For the project, the student must design, train and validate a deep learning approach for image or text analysis and the deliverable is the result of this learning procedure and a written report describing the motivation of the solution, the technical details of the method, the experimental results and the conclusions. It aims to verify the applied knowledge and skills of the student. The oral interview has an approximate duration of 45 minutes, and includes the discussion of the project work, aimed at verifying the student’s ability to describe the solution, to defend the design choices and to discuss the pros and cons of the proposed approach. The oral interview will also verify the theoretical knowledge on all the topics of the course.

Verification of learning

The exam consists in a project and an oral interview. For the project, the student must design, train and validate a deep learning approach for image or text analysis and the deliverable is the result of this learning procedure and a written report describing the motivation of the solution, the technical details of the method, the experimental results and the conclusions. It aims to verify the applied knowledge and skills of the student. The oral interview has an approximate duration of 45 minutes, and includes the discussion of the project work, aimed at verifying the student’s ability to describe the solution, to defend the design choices and to discuss the pros and cons of the proposed approach. The oral interview will also verify the theoretical knowledge on all the topics of the course.

	Texts
	Zhang, A., Lipton, Z. C., Li, M., & Smola, A. J. (2021). Dive into deep learning. arXiv preprint arXiv:2106.11342. https://d2l.ai/ Supplementary teaching material will be available on the e-learning platform (http://elearning.unisa.it) accessible to the students using their own credentials

BETA VERSION Data source ESSE3 [Ultima Sincronizzazione: 2024-12-13]

Mario VENTO | ADVANCED MACHINE LEARNING