Fundamentals of Deep Learning

What is deep learning? When to use it? How to use it? Why use it and why not use it? Where is the field now, and where is it going towards?

The course will take place over three days, five hours every day. Of those five hours, three will be theoretical content, and two will be hands-on experimentation.

• Dario Garcia Gasulla
• Anna Arias Duart

• Dario Garcia-Gasulla is an AI researcher specializing in deep learning, explainable AI, and interdisciplinary collaboration. His work brings together AI research with real-world impact, in the context of and HPC. He founded the HPAI group at BSC, where he contributed in interpretability, scalability, and ethical deployment of AI systems, and has worked in a variety of domains. His past experience includes a dozen EU funded projects and seven years as a lecturer on AI. He now acts as Research Area Director at the AI institute of BSC’s, where he supports multiple teams and manages specific research lines on different applied topics.

• Anna Arias Duart is the group leader of TruE, the Trustworthy and Evaluation research group at BSC. Her work has focused particularly in enhancing model evaluation, explainability and model safety, making multiple contributions in those fields for both image and text-based models.

This course starts from the most basic components of deep learning (neuron, activation function, loss), and builds up to the most popular architectures (CNNs, LSTMs, Transformers). It reviews the contributions that have surpassed the test of time (residuals, bottlenecks, attention, etc) and highlights the strong points and weaknesses of these methods. Good practices and common tricks of the field are reviewed. In the practical side, attendants will experiment with the training and evaluation of different model architectures for three different problems.

Although the course has very little formal content, basic math and engineering skills are needed to follow the concepts being introduced. Fundamentals of machine learning (what is a learning task, what is a dataset, etc.) are necessary. For the practical part, knowledge of Python and familiarity with a command line (cd, cp, ssh, scp, etc.) are needed.

Certificate of Achievement

The labs will be done in a combination of online resources (colab) and hosted HPC infrastructure. No local installation or software will be needed. For HPC access, a command line interface is required.