Deep Learning

Lecturer Prof. Dr. Andreas Zell
Office Hours Wed. 13:30 - 15:00
Dates/Time Wed. 10:15 - 12:00
Volume 3V+1Ü (6 LP)
Start Wed. 20.10.2021
Location Morgenstelle, Hörsaal N5
Stream

In parallel, the lecture is streamed (and recorded) via Zoom:

https://zoom.us/j/93601758187?pwd=TlhGOXNrQjlCTVhkSTVEeTdBUDBpUT09
Meeting-ID: 936 0175 8187
Kenncode: 808149

Cycle annually during the winter semester
Klausur probably Wed. 16.02.2022, 10:15 - 12:00

In Alma-Vorlesungsverzeichnis

Description:

The lecture will cover the main principles and models of "deep neural networks", adapted from the textbook "Deep Learning" by Goodfellow, Bengio and Courville. These network models are currently top performers in recognition for many machine learning problems, from object recognition, image segmentation, speech recognition, handwriting recognition, to time series analysis of financial data. They are used by nearly all car manufacturers for driver assistance systems and autonomous driving, by Internet companies for semantic labeling of image data, and in mobile assistants such as Siri, Cortana, and Google Now.

After an introduction to the topic, mathematical basics will be covered first: Linear Algebra, Probability, Numerical Foundations, and Fundamentals of Machine Learning. Then, deep neural networks and techniques such as backpropagation are covered; followed by regularization techniques for deep neural networks (which improve their generalization performance on unknown test data). This is followed by optimization techniques for training deep neural networks and a chapter on so-called Convolutional Neural Networks (CNNs). Further chapters on recurrent networks, methodology in practice and an application chapter are planned. At the end, specific architectures of deep neural networks will be discussed, e.g. AlexNet, GoogleNet, Yolo, Residual Networks, DenseNet.

The lecture covers about half of the 800-page textbook and follows it very closely.

In the exercise, the theoretical knowledge acquired in the lecture will also be deepened by solving practical tasks. In addition to programming tasks in Python, the deep learning library PyToch from Facebook will be used.

Requirements:

The lecture is intended for master students!
Good knowledge of mathematics (at least "good" in mathematics I - III) is required.
The material is very extensive and contains many formulas (and too much text on the slides).
Python programming skills are assumed.

Registration (Mandatory):

Via ILIAS: Deep Learning (ML4103)

Literatur:

  • Lecture notes (will be made available as PDF in Ilias before the lecture)
  • Ian Goodfellow, Yoshua Bengio, Aaron de Courville: Deep Learning, MIT Press, 2016.The book is available online on the WWW, cf.
  • http://www.deeplearningbook.org/

Exercises for the lecture Deep Neural Networks

Teaching Assistants: Valentin Bolz, Timon Höfer, Maximus Mutschler, Andreas Ziegler
Office hours: usually in the exercise, in exceptional cases by arrangement
Date/time: Wed. 12:15 - 14:00 (sometimes first hour of the exercise is used for the lecture as well)
Place: Morgenstelle, Hörsaal N5
Stream:

In parallel, the session is streamed via Zoom:

https://zoom.us/j/91793488840?pwd=SlZsQmM0bFY3dWpxWXpDVXVrY2J6dz09
Meeting-ID: 917 9348 8840
Kenncode: 020803

Issue of the exercise sheets: During the lecture
Submission of the completed exercise sheets: none, solutions will be presented one week after issue