CS 559 Deep Learning (Spring 2017)

Instructor: Gokberk Cinbis (Room: EA-431)
Office hours:  Mon 10:40-11:30
Syllabus:  Please read the syllabus for course details.

Syllabus

Week Lectures and presentations Resources
W1a

Feb 8

Lecture:
  • Introduction, course logistics
Part I of Deep Learning by Goodfellow et al.
W1b

Feb 10

Lecture:
  • Hallmarks of deep learning
  • An image classifier example
Python is strongly recommended for the projects and homeworks: an excellent tutorial by Justin Johnson.
W2a

Feb 15

Lecture:
  • Linear classification and loss functions
Lecture notes by Andrej Karpathy: loss functions
W2b

Feb 17

Lecture:
  • Linear classification and loss functions (cont'd)
  • Optimization
  • Backpropagation
Lecture notes by Andrej Karpathy: optimization
W3a

Feb 22

Lecture:
  • Backpropagation (cont'd)
W3b

Feb 24

Lecture:
  • Feedforward networks, activation functions
Section 2.6-2.8 of Deep Learning by Goodfellow et al. [optional]
W4a

Mar 1

Lecture:
  • Basics of neural network training
W4b

Mar 3

Lecture:
  • SGD variants, dropout regularization
W5a

Mar 8

Lecture:
  • Convolutional neural networks - basics
W5b

Mar 10

Paper presentations:
W6a

Mar 15

Lecture:
  • Convolutional neural networks - architectures
W6b

Mar 17

Paper presentations:
W7a

Mar 22 (+1h)

Lecture:
  • Deep learning for spatial localization
  • Visualization of ConvNets
W7b

Mar 24

No class.
W8a

Mar 29

Lecture:
  • Visualization of ConvNets (cont'd)
W8b

Mar 31

Paper presentations:
W9a

Apr 5 (+1h)

Lecture:
  • Recurrent neural networks
W9b

Apr 7

Midterm:
  • April 7, 2017, 10:40-12:30, EB-204 (in-class).
  • Topics: weeks 1-8 (inclusive).
  • You may bring only course slide printouts.
W10a

Apr 12

Lecture:
  • Word embeddings and language models
W10b

Apr 14

Paper presentations:
W11a

Apr 19

Lecture:
  • Unsupervised learning and deep generative models
Tutorial on Variational Autoencoders
W11b

Apr 21

Project progress presentations
W12a

Apr 26

Lecture:
W12b

Apr 28

Paper presentations:
W13a

May 3

Paper presentations:
W13b

May 5

Paper presentations:
W14a

May 10

Paper presentations:
W14b

May 12

Project final presentations

Homework

Super-resolution in TensorFlow. Due: April 1st, 2017, 23:50.

Paper presentation guidelines

  • Each presentation should take around 22 minutes, plus 2-3 minutes of discussion.
  • Please prepare your presentation as a short-lecture that focuses on a particular work, rather than a dry summary of the contents of the paper.
  • Therefore, please cover important related work (if not already throughly covered in the class) in your presentation, in order to (i) make the presentation accessible for everyone, and, (ii) properly discuss the strengths and weaknesses of the paper compared to related work.
  • Each presenter is required to send me a complete draft of the slides (as a pdf) 2 days before the presentation date, and the final slides before class on the presentation date. Slides will be published on the course webpage.
  • Please also see Week 1 slides for some additional information.

Midterm

Midterm is scheduled for April 7, 2017, 10:40-12:30, EB-204 (in-class).

Project

Projects can be done individually or as a group. The goal of the projects will be to explore novel applications of contemporary deep learning techniques or develop novel deep learning techniques. Projects related to research topics of the students are encouraged.

  • Project groups Each project group is required to send the names of their members. Due: March 10, 2017, 23:59. [Groups]
  • Project proposal Each project group is required to send a one-page project proposal. Due: March 24, 2017, 23:59.
  • Progress report Each group will prepare a progress report and presentation. Due: April 18, 2017, 23:59.
  • Final report Each group will prepare a final report and presentation. Report due: May 24, 2017, 23:55.
  • Please use the CS559 Moodle submission page. Emailed submissions (unless Moodle is down) and late submissions will not be accepted.
  • All reports must be prepared using the IEEE double-column conference template. Using LaTeX (or LyX) is recommended. Final report should be around 8 page long.
We reserve the right to make changes in the course content.