91250 - Deep Learning

Academic Year 2019/2020

Learning outcomes

At the end of the course, the student understands the foundational ideas, recent advances and application potential of deep neural systems. The student understands supervised and unsupervised techniques, basic neural topologies, methods for visualizing and understanding the behavior on neural nets, adversarial and generative techniques, reinforcement learning, and recurrent networks. The student is able to apply such technologies to solving classification problems in realistic domains.

Course contents

The course starts with an introduction to Deep Learning, Neural Networks and their typical learning mechanism: the backpropagation algorithm.

We shall discuss the main types of neural nets: feed forward,
convolutional and recurrent, providing concrete examples and discussing successful neural architectures for image processing, localization, segmentation, style transfer, text processing and many other tasks.

We shall investigate techniques to visualize the behavior of hidden units (tightly related to deep dreams and inceptionism), techniques to fool neural networks, and modern generative approaches comprising Variational Autoencoders, Generative Adversarial Networks, and their most recent developments.

The final part of the course will be devoted to an introduction to Deep Reinforcement Learning, with applications to the design of intelligent agents for video games, autonomous driving and other situations requiring complex and adaptive behaviors.

Readings/Bibliography

Suggested reading:

I.Goodfellow, Y.Bengio, A.Courville Deep Learning MIT Press.

 

Specific pointers to on line material will be provided at each lesson, in addition to the slides of the course.

Teaching methods

Frontal lessons based on slides, with discussion of practical examples via pyhton notebooks.

Assessment methods

Home project on a restricted set of topics selected by the teacher.

Projects can be done in teams of up to two members. The exam will consist in the presentation and discussion of the homework. 

The assesment method may change in relation with the attendance.

Teaching tools

Lectures will make extensive usage of slides. Working examples will be delivered by means on python notebooks.

Office hours

See the website of Andrea Asperti