“AI – Advanced Level” – in-person course in three modules

12-13-14 March at MUG
Via Emilia Levante 9/f (BO)

Executive summary

The course aims to provide an overview of the world of Deep Learning (DL), which is increasingly linked to supercomputers given the growing amount of data and need for computing power. The first module is designed to offer a general overview of the main neural network architectures for various applications, including image classification, complementing theoretical explanations with practical exercise sessions to make participants autonomous in using such algorithms. The second module will provide the basic knowledge for using an HPC machine to understand its potential and necessity in training networks with large amounts of data. Then, some use cases will be provided to motivate the use of HPC in the DL field, such as large networks that can be adapted to specific use cases. In a practical session, some DL models will be implemented on HPC infrastructure using parallelization techniques. The last module will be dedicated to exploring different architectures for time series analysis and fine-tuning Large Language Models for specific use cases.

Topics

  • Main data-related issues
  • Main neural network architectures
  • Basic concepts of HPC
  • Parallelization of neural networks
  • Fine-tuning of neural networks
  • Neural networks for time series analysis
  • Large Language Models

Course Participants

Individuals with a technical-scientific background, basic programming language skills, familiarity with the Linux operating system, and basic concepts of Machine Learning, interested in seeing and trying advanced Deep Learning techniques using HPC.

Lecture Days

  • 12/03/2024 Time 9:00-13:00
  • 12/03/2024 Time 14:00-17:00
  • 13/03/2024 Time 9:00-13:00
  • 13/03/2024 Time 14:00-17:00
  • 14/03/2024 Time 9:00-13:00
  • 14/03/2024 Time 14:00-17:00

Mode of Attendance

The Course consists of 3 in-person days.

Lecturers

Has earned a PhD in Mathematics with a thesis on tensor and matrix decompositions for image classification. Currently works in the HPC Department at CINECA, focusing on the use of HPC technologies in artificial intelligence projects.

Recently graduated in Mathematics with a master’s thesis focused on data compression techniques applied to the field of Digital Twins. Currently engaged in an internship at the HPC Department of CINECA, where she is finalizing her studies on compression and contributes to the development of HPC technologies within artificial intelligence projects.

Earned a PhD in Computer Engineering at the University of Modena and Reggio Emilia. Currently a Data Scientist and AI consultant at NVIDIA, where he also oversees the NVIDIA AI Technology Center for Italy, a collaboration between NVIDIA, CINI, and CINECA to accelerate academic research in the field of artificial intelligence. For over 14 years, he worked as an HPC specialist at CINECA, providing support for large-scale data processing.

Earned a Master’s degree in Artificial Intelligence, presenting a research thesis in computer vision on deep learning techniques for monocular depth estimation on high-resolution images. Currently collaborates with CINECA and Leonardo SpA, focusing on the development of artificial intelligence technologies in the HPC sector.

Earned a Master’s in High Performance Computing at SISSA in Trieste and worked as an HPC researcher at OGS in the field of oceanography. Subsequently started working at CINECA, where today he serves as a consultant for industrial applications and technical project manager for technology transfer projects towards industries, especially SMEs.

Andrea Pilzer is a Solution Architect at NVIDIA within the NVIDIA AI Technology Center in Italy. He was a postdoc at Aalto University, working on uncertainty estimation for deep learning. He has worked at Huawei Ireland and earned his PhD in Computer Science at the University of Trento with Prof. Nicu Sebe and Prof. Elisa Ricci.

Earned a Master’s degree in Artificial Intelligence with a research thesis on the application of Graph Neural Networks in the NLP field to optimize sentence classification performance. Currently collaborates with CINECA and Leonardo SpA, focusing on the development of artificial intelligence technologies in the HPC sector.

Course program

Time: 9:00-17:00

Objectives:
The first module will be dedicated to introducing concepts related to deep learning and the main architectures of neural networks, as well as to implementing the described models in Python language.

Contents:
Introduction to the main issues related to data
Description of the architecture of some neural network models
Python language exercise on the described models.

Time: 9:00-17:00

Objectives:
In the first part, some basic knowledge in the HPC field will be provided. In a practical session, it will be shown how to interact with a supercomputer (login, data upload, code execution).

In the second part, some use cases will be described to motivate the use of HPC in the DL field.

In a practical session, it will be shown how to parallelize neural networks on HPC infrastructure and how to perform fine-tuning to adapt a pre-trained network to one’s use case.

Contents:
Introduction to HPC
HPC environment exercise
Exercise on parallelization of deep learning models in an HPC environment
Fine-tuning of neural networks.

Time: 9:00-17:00

Objectives:
The first part will be dedicated to presenting various neural network architectures for time series analysis (e.g., anomaly detection). The second part will focus on language models, analyzing their architecture and showing the workflow for fine-tuning an LLM.

Contents:
Time series analysis with various neural network architectures
Introduction to LLM models
Fine-tuning of LLM.