Deep Learning, 5+3 credits

RECOMMENDED PREREQUISITES

Basic undergraduate courses in linear algebra, statistics, probability, optimization and programming experience in Python, MATLAB or similar.

LEARNING OUTCOMES

After passing the course the student should be able to:
- describe and use backpropagation together with gradient descent and stochastic gradient descent to optimize a model
- implement a fully interconnected multi-layer neural network
- explain under and over fitting and what can be done to avoid them
- describe and use different kinds of regularization techniques
- describe and use deep convolutional networks for classification and regression
- describe and use deep learning models for timeseries data
- use modern environments for deep machine learning to solve practical problems

LEARNING OUTCOMES FOR DOCTORAL DEGREE 

A1 – Data driven machine learning in general and deep learning in particular is becoming and important knowledge for virtually any scientific field where big about of data needs to be processed and analyzed.
B5 – Deep learning is a field which moves tremendously fast. We will in this course only be able to scratch on the surface, but we aim to give the students tools to identify the need for knowledge within this area.

For those who also opt for taking the additional project course we also address:
B2- The student formulate their own project and plan for the appropriate methods within a given time frame.
B4- The PhD student present and discuss the work from the project both in speech and writing.

COURSE CONTENTS

The lectures will be focused on theoretical aspects and the mandatory hand-in assignments on practical implementation of deep learning methods. The course will therefore deal less with particular applications within the field.
Feed forward neural networks, backpropagation, stochastic gradient descent, convolutional neural networks, residual neural network, semantic segmentation, instance segmentation, over-/underfitting, bias-variance, regularization, practical methodology, batch normalization, transfer learning, different deep learning models for timeseries data.

INSTRUCTION

Lecture series: 8-10 lectures, (2 hours each). The lectures are given by the teachers.
Helpdesks: Sessions where students can get help with the hand-in assignments (see below).
Project (optional): Successful projects will be awarded an additional 3 hp. This is a great mechanism to spark bigger projects and spin-off collaborations.

The lectures and helpdesks will be scheduled into three two-day blocks to facilitate for students travelling from other universities/campuses.

ASSESSMENT

3 to 4 bigger hand-in assignments. All assignments will be focused on implementation aspects for deep learning algorithms and models. In the first exercises, standard deep learning methods will be implemented from scratch, and in the last exercises a state-of-the-art high-level software library for deep learning will be used. 

COURSE EXAMINER

Niklas Wahlström, niklas.wahlstrom@it.uu.se

DEPARTMENT WITH MAIN RESPONSIBILITY

Information Technology

CONTACT PERSON/S 

Niklas Wahlström, niklas.wahlstrom@it.uu.se

APPLICATION

Submit the application for admission to:
 https://doit.medfarm.uu.se/bin/kurt3/kurt/63486

Submit the application not later than: February 1, 2023, but the first come, first served principle applies in general. To complete the application a successful submission of a pre-course assignment needed no later than March 1, 2023.
See more information about the pre-course assignment and the course in general on the course homepage https://uppsala.instructure.com/courses/72437